Release history

The activation window shows licence status, client, email, edition, expiry, machine ID with Copy ID, activated/allowed computers and the last online check. Computer counts reflect that check, not continuous monitoring.

Build 16 — DEMO mode and full-version activation

From build 16, the licence bar shows DEMO or FULL VERSION. Editing, previews and project saving remain available without a licence. Projects retain every source and their full duration; only delivery copies are limited.

DEMO video and batch exports are limited to 60 seconds, a maximum of 1280 × 720 and 30 fps, with the StreamerEditorPro DEMO watermark. Aspect ratio is preserved; portrait video fits within these bounds. Audio exports and AI narration are limited to 60 seconds. Delivered demo AI images and YouTube cards are limited to 1280 × 720 and watermarked. AI video variants follow the same video limits. AI previews remain available; optional local components require separate installation.

Click Activate full version, paste the key issued for StreamerEditorPro and click Activate with Internet access. Once validated, new exports immediately use the full features without reinstalling or reopening the project. An export already in progress keeps its starting mode. Existing demo files remain watermarked and can be exported again from the complete project.

Activation is bound to the computer and the authorised computer limit. A StreamerPro key does not unlock StreamerEditorPro. Checks run at startup and periodically; a signed activation may work offline for up to 30 days, never beyond licence expiry. Network downtime does not erase a still-valid activation. Suspension, revocation and released computers take effect at the next online check; the offline allowance is the maximum delay. Check online provides a manual check.

The licence stays in the Windows account’s user data, outside projects and the installation folder. Updates preserve this data. A computer change may require releasing the old computer in Administration → Licences StreamerEditorPro. Administrators create keys, choose duration and computer limits, suspend or revoke licences, release computers and inspect history. Statistics distinguish StreamerEditorPro and mark incomplete periods. No automatic Stripe payment flow is added for this product.

• Installation assistant: open Tools → Installation assistant, Updates → Installation assistant, or Download components in a local AI window. The same Windows assistant is available on streamerpro.fr. Choose the latest tested release or the stable reference of 8 November 2026, then Windows installation or Portable. Portable extraction uses a separate folder for each version and build. Shared AI image/video engines, voices in all five languages and local AI text are optional and separate from the application. Full offline archives remain available. Photo styles need no AI model. Progress uses actual bytes for each stage, speed and estimated time remaining. Pause or close preserves partial files; reopen, select the same items and choose Resume/Download and prepare. Interrupted connections resume in the same way. The assistant verifies full hashes before use, reuses identical installed components and restores previous files if a component installation fails. Insufficient space blocks the operation. Starting Windows installation asks for confirmation, backs up settings and asks to save the open project. Stable rollback reinstalls the reference application; it does not convert newer projects or delete media. Displayed size is a maximum before reusable components are detected.

• Shared video and thumbnail styles. In Automatic with local AI, the selector provides the same 40 looks: 31 photo styles that preserve the scene and 9 artistic AI styles that may reinterpret objects and geometry. Follow the last style selected in the video generator is enabled by default. The last video choice is remembered by language-independent ID. If no video style has been saved, Natural cinema is used, matching the video generator. Selecting another thumbnail style disables following; enabling it again restores following. This choice is saved with the project and also applies to local bookends created during analysis and automatic montage.

• Reworked video styles: explicit distinction between 31 source-preserving photo styles and 9 material-transforming AI styles. Individual palettes; removed the RGB blend that homogenized colors. AI keyframe styling and source-motion propagation with matching preview. Irrelevant controls hidden by mode; offline components unchanged.

• Local AI video: automatic source-frame before/after preview using the same video model, obsolete calculation cancellation, memory cache and source fidelity. Window shown after construction; size, position and maximized state saved automatically.

• Local AI video. 40 translated styles, including 37 new presets, reusing existing models without changing the offline pack. Stage percentages use actual FFmpeg counters and AI iterations; final 100% follows validation, with explicit cancellation and error states.

• Timeline controls. Close moved to the bottom-right bar shared by all six tabs; Save and Synchronize cameras with music reordered. The three main Timeline commands are aligned in one row.

• Visual timeline. Help text fills the frame width and the tools panel fits the active tab, width and language, reducing empty areas while preserving access to tracks.

• Local AI video. Added source-conditioned Wan VACE generation with short variants, styles, optional instructions, embedded preview and explicit undoable timeline insertion. Offline pack now includes video models; interface, manual and history in five languages.

• 3D Logos. Replaced all 11 sheets with large-format versions and added 600 logos across 20 new themes. Themes use a list, with translated names and loading limited to the open theme.

• The Streaming category contains 30 badges. Choose 4K Ultra HD, 24/7 Live, Alert bell, Drone 1080p, Drone 4K/UHD or Like in one of five color series.

• 3D Logos. Thumbnails and the scrollbar appear when a library tab first opens, including 3D Flags and 3D Arrows.

• The shared library contains 181 logos, 60 flags and 60 arrows. Open 3D Arrows, then Multicolor or Gold, to choose a variant. Add these assets to thumbnails or overlays.

• The 3D Flags tab contains all 30 flags from the sheet with their original colors and transparency. Add them to thumbnails or overlays from the same library.

• In Edit title, text and logos and in Overlays, open 3D Logos to choose from the 30 sheet logos and the Subscribe button. In Overlays, confirm the logo properties to add it to the edit. The thumbnail preview is centered in the editor.

• Thumbnails. The six Smile, Glasses, Heart, Star, Arrow and Shocked stickers now have an editable 3D appearance.

• Start and end thumbnails now show a large preview of the selection on the right. Layer width in the editor can reach 1,000%.

• Postcards add six selectable layouts, an automatic variety mode and automatic style mixing. Two-photo layouts use distinct frames from the edit. Graphic elements remain editable.

• Local AI: the Postcard template creates six proposals at project resolution. Shared details are optional and captions use the available context of each automatic capture. All text remains editable.

• The Opening and closing thumbnails window displays its controls while the gallery loads in the background. Cached images are reused and missing images are restored from the project. You can close the window during loading.

• The thumbnail choice is stored as soon as you click. The log reports the selected mode when analysis starts and confirms insertion at the end. The gallery identifies thumbnails created with local AI.

• Automatic Montage now defaults to opening and closing thumbnails with local AI for new projects. It inserts two five-second thumbnails. The options to use frames without AI or skip thumbnails are available.

• Local AI thumbnails now offer a YouTube style with stronger color and contrast. Three opening and three closing proposals match the project resolution. Titles, text and logos are editable layers, and edits are saved as variants.

• Local AI thumbnails use the framing and details of the captured video frame with a restrained color and lighting adjustment. The PNG dimensions match the project or Short. Automatic selection favors sharp, distinct frames. Three opening and three closing proposals are available.

• StreamerAI Studio provides default local AI thumbnail generation in the opening and closing assistant and in Shorts Studio cards, plus WAV narration in five languages from Audio. The 64-bit Windows pack containing the engines, models and voices is a separate download from streamerpro.fr.

• Shortcuts after Next / Previous: editor commands are reattached to current controls after each scene load and when the reader is ready. Common commands run before local field handlers to prevent interception. Focus is restored after layout without forcing another application to activate or leaving an active input field or child dialog. Callbacks from an old preview are ignored. Ctrl+1, Ctrl+2, Ctrl+4, Ctrl+8 and Ctrl+0 are keyboard keys rather than mouse buttons. They retain their temporary preview-speed role with settings hidden and in F11.

• Edit scene shortcuts: Ctrl+U triggers Move up and Ctrl+D triggers Move down, as in the Timeline tab. They use the same buttons, respect their boundaries and work from fields, with settings hidden and in F11 fullscreen. Tooltips show the shortcuts. Child dialogs do not forward them to the editor. Ctrl+Shift+D does not trigger Move down.

• Edit scene: Move up and Move down are immediately to the left of Save, with matching alignment. They move the displayed scene one position in the montage using the same command as the Timeline tab. Moves take effect immediately and are recorded in timeline undo history; Apply is not required. The same scene remains displayed and selected, current fields are retained and the reader is not recreated. Next and Previous then follow the new order of scenes open in the editor. The commands remain available with settings hidden and in F11, and are disabled at montage boundaries. Moving a scene does not change its duration or speed.

• Navigation in Edit scene: on opening and with Next or Previous, the timeline selection follows the displayed scene. The matching row is brought into view in the main list; if the visual timeline is open, the matching clip is selected. Matching uses the scene ID, including multiple scenes from one file and after deletion or restoration. The editor’s initial selection for batch operations is retained. Following the scene does not move keyboard focus or change playback/pause intent.

• Edit scene: Integer speed-up (label and ×1 to ×100 selector) is in the preview speed bar, between Edit speed and Advanced options / Hide advanced options. Controls are aligned on the same row. The setting remains accessible with scene settings or advanced options hidden, including F11 fullscreen. It changes the scene speed; ×1, ×2, ×4, ×8 and ×10 retain their temporary preview-speed role.

• Initial automatic montage settings: Quality defaults to 0.5 and HDR Effect is enabled, with strength 20. Resetting automatic montage settings restores these values. Values already saved in a project are retained when opening it. HDR Effect is an image effect; it does not force HDR10 export.

• Delivery identity: new packages show the version and a build number in windows, About and filenames (example: v0.20.368_build0002). Each build reserves a distinct number, including failed builds. output/lots keeps both packages and lot.json with checksums and test evidence; existing lots cannot be overwritten. In Updates, Tests for this build shows only results matching the current build. Choose stable delivery verifies and selects a saved lot.json; Restore stable delivery asks for confirmation, verifies files again and starts Setup after the usual project save prompt. Preferences are backed up before launch. The user selected the 8 November 2026 delivery (887 tests across 114 modules) as the stable reference; its old executables have no embedded build number. Numbered server preparations target beta and do not replace stable automatically. Restoring an older app does not convert projects saved in a newer format: retain a compatible project copy.

• First audio playback in the visual timeline: the playhead can be drawn before the initial layout finishes without a “timeline_height” error. Its height then follows the displayed tracks.

• Render cancellation: the progress window closes as soon as the task confirms it has stopped, and the status shows “Render cancelled”. It no longer remains on “Cancelling…” after processing ends. The elapsed time shown covers the entire render. Valid completed blocks remain available to resume an identical export.

Speed and playback during an Edit scene session: the ×1, ×2, ×4, ×8 or ×10 choice is retained with Next and Previous. If playback was active, it resumes on the chosen scene at that speed in windowed, maximized and F11 fullscreen modes; a paused preview stays paused. The request waits for metadata and player opening, including file changes or player recreation. Pause or Stop cancels the wait; closing, an opening error or a timeout must not cause delayed playback. Media access may still create a loading interruption. Montage speed explicitly restores each scene’s own speed. A new window opens paused at montage speed. Saved scene and export speeds remain unchanged.

After deleting in Edit scene: Quit or the window close control also offers “Fit to music duration”, even if you only deleted videos without clicking Apply. One or more deletions still in effect are included alongside applied duration changes. A deletion restored with Ctrl+Z no longer counts. The proposal requires active music and at least one remaining video. Declining closes the window and keeps the deletions; accepting opens the usual timeline proposal after the editor closes, without automatically confirming the adaptation.

Fluid-passage shortcuts in Edit scene: Ctrl+Shift+F recalculates the fluid passage, Ctrl+Shift+N uses the next passage and Ctrl+Shift+P the previous one. They run the same commands as the buttons, with the same locks and limits. They work from settings fields, with settings hidden or during fullscreen preview. A disabled command or an ongoing recalculation does not trigger another action. Translated tooltips show the shortcuts. Ctrl+N and Ctrl+P still switch scenes; Ctrl+Shift+N remains New project in the main window.

Save and Ctrl+S save to the active file. A new project prompts for a file on its first save. Save As and Ctrl+Shift+S open the name and folder chooser with the current project values. The selected file becomes active after a successful save; cancel keeps the current path.

Apply and Save in projects with network media: only changed scenes are sent to the project. End-bound validation uses loaded durations and a media index, without looking up NAS files on the interface thread. An invalid duration prevents the whole set of edits from being committed; applying again without changes does not rebuild the timeline. This validation does not replace source availability checks when media is read.

Montage construction and NAS sources: passage grouping, folder assignment and quality control compare recorded paths without repeatedly resolving each network file. Paths are still checked when operations actually read media. Selection progress shows the number of scenes processed; Cancel is checked between scenes and while grouping candidates. The log separates source indexing, selection, audio and quality control, camera/drone groups and opening/closing card creation. These measurements help locate remaining waits separately from preview encoding.

Preparation while switching scenes: a short 0.6-second settling period precedes a new uncached encode and preparation of other sources. Ready caches remain available without this delay. A running GPU encode may continue during rapid navigation or a return to a completed cache, avoiding unnecessary restarts. If a new uncached scene remains selected, its preparation takes priority and may interrupt the old job to release the GPU; that job restarts from the beginning with the same GPU method when resumed. GPU encodes remain serialized. CPU jobs retain suspend/resume. Closing or disabling preparation still stops or pauses jobs as before.

GPU analysis and 10-bit video: CUDA decoding and scaling can remain on the GPU; the small resized image is then converted to grayscale for analysis. This avoids an observed FFmpeg crash during luma extraction after a GPU transfer. A pipeline is confirmed only after successful processing. If hardware scaling fails, diagnostics retain the precise error, including an exit code without a message, even when hardware decoding with CPU scaling succeeds. Maximum GPU still prioritizes a validated GPU, and Strict GPU still forbids CPU fallback.

SEP icons and Windows scaling: the logo keeps its artwork and colours. The ICO now contains 26 native sizes from 16 to 256 pixels, including 20 and 40 pixels for 125% scaling. Windows select their small and large icons for the current monitor and reapply them after creation or a monitor change. This covers the main window, secondary windows, manual, history and GPU/HDR previews. GPU windows also receive their native icon after their window is created and after a display-mode change.

Secondary windows share icon images and reuse native resources. Closely spaced size changes are coalesced, and deferred updates are cancelled when a window closes. This correction does not change overall application scaling. The executable, shortcuts and installer use the same multi-resolution ICO on their next rebuild; an already open application must be restarted with the updated build to display the correction.

GPU preview size: new Edit scene caches automatically adapt their bitrate to the full video duration. NVIDIA NVENC, Intel QSV and AMD AMF use a calculated target bitrate, maximum bitrate and buffer to leave headroom below the 2 GB per-file limit. A 25% reserve covers encoding variation and container overhead; longer videos use a lower bitrate. Frames remain independent to preserve fast forward/backward navigation.

If the file still exceeds the limit, at most two retries keep the same GPU engine with a lower bitrate and maximum dimensions of 960×540, then 640×360. The first attempt remains limited to 1280×720. Actual dimensions are displayed when height is below 720 pixels. A size error alone does not switch GPUs or trigger CPU fallback; it is separate from the five attempts for a hardware failure. File-size and free-disk-space limits remain active.

These settings affect preview files only: sources, montage, audio and final exports remain unchanged. Valid existing caches are reused. Logs record duration, bitrates, size budget, maximum dimensions and size-related retries.

Reopening previews with NAS media: background preparation checks the cache before probing the original video. A valid local cache supplies its duration directly, without running FFprobe on the network media again. Source identity is still checked; a missing, corrupt or duration-less cache still requires source metadata. Logs include the source, cache path, reuse time and, when necessary, metadata-reading time. Cancelled or failed preparation is recorded explicitly. These measurements distinguish reopening a cache from creating a new preview and do not guarantee identical speeds on every computer.

Identifying older 720p caches

Older caches are also identified when their video stream contains the FFmpeg encoder signature: NVENC, QSV, AMF or CPU libx264/MJPEG. Without a usable signature, the original encoder remains unconfirmed; this message does not indicate a GPU error or trigger recalculation. The status distinguishes the displayed scene’s cache from preparation of other sources in the background: the file count and other source’s percentage do not mean that the ready cache is being rebuilt.

720p previews — GPU priority and retries

Preparation keeps the GPU engine verified at startup and tries its variants before another GPU. After a failure, each GPU method can be attempted up to five times in total, with a short cancellable delay between attempts. An explicit driver incompatibility or unsupported command proceeds directly to the next method. A secondary engine must pass its own test before being offered; its presence in FFmpeg is insufficient. AMD AMF can also encode previews, with CPU decoding and scaling. CPU fallback remains available after the available GPU methods fail.

When an unfinished preparation is put on hold to change scenes, its GPU process is stopped to release resources. It restarts from the beginning with the same GPU method when reactivated; completed previews remain reusable. GPU preparations from different windows run one at a time. These changes concern preview creation; the player selects hardware decoding separately.

Retries appear as “Retrying NVIDIA NVENC (2/5)…”, without presenting a recovered failure as a current fault. Detailed errors remain in diagnostics. After successful creation, the status confirms the encoder used and clears the previous fallback reason. New caches store their creation encoder; valid older caches remain reusable and show “original encoder unconfirmed” if this information is missing. H.264 is a video format and cannot identify the GPU on its own.

Edit scene: Next and Previous immediately hide the image retained by the player. The new still appears as soon as it is ready; the video surface is revealed only after the first frame at the new playback position is confirmed. This also covers two passages from the same file and rapid navigation without rebuilding the player for every scene.

720p previews: variants of the GPU engine confirmed at startup are tried before another GPU. A QSV decoding or scaling failure alone no longer invalidates the QSV encoder. Player positioning, first-frame and network errors no longer trigger automatic CPU fallback without an explicit hardware failure indication. Interpolation keeps CPU calculation but now tries NVENC or QSV encoding, using GPU decoding followed by a CPU-decoding variant with the same encoder. CPU encoding fallbacks remain available after GPU paths fail. Status messages and logs distinguish CPU interpolation, encoder and fallback reason. Manual updated in five languages.

Persistent smooth previews: saved projects use an adjacent <project name>.previews folder, reused automatically when reopened. Compatible local previews are recovered in the background, including after a project’s first save. Files remain available despite temporary preview cleanup on exit. Source, interpolation settings and readability checks prevent outdated previews from being reused. Copies are published only after completion. The manual explains moving a project with its folder and keeping previews on SSD for NAS sources.

Edit scene — display and pause: a still frame is prepared in the background from a valid local preview when available, or from the source video on the NAS. The video player waits for a click on Play; opening the window and moving the cursor no longer trigger temporary playback. Pause keeps the current media and displayed frame without automatically reopening the NAS original. A pause request also cancels positioning or resuming still pending. Stop returns to the start of the passage using a still frame. The duration saved in the project makes the controls available before its background verification. Command states and styles are updated only when they change. The 720p preview option now covers playback and navigation; disable it to work directly from the originals. These preview changes do not modify the montage or source files.

NAS verification: two 4K HEVC 10-bit source videos were used on this PC for 21 Play/Pause cycles, with and without a local preview. Pause remained stable once established, and playback did not start before Play. Without a local preview, the player sometimes needed about 1.5 seconds to pause. These measurements apply to the tested PC and do not constitute validation on the ROG laptop.

Edit scene: opening the window and using Next/Previous compare media paths without querying the NAS from the interface. Preview progress no longer checks files while holding the lock shared with the window. Saved fluidity zones are displayed and then verified in the background; valid zones with existing analyzed passages avoid another analysis. Duration metadata are reused for 30 seconds when returning to a source. Each request has its own identifier: navigating before the previous read finishes no longer causes a task conflict, and cancelled results are ignored. Reused previews are revalidated on their reader task; a changed source is rejected. The log records interface preparation and metadata timings. The first image still depends on NAS access and decoding.

Opening projects with NAS media: reading the SEP file, checking its data and restoring cards and AI audio tracks run in the background. The status bar shows the stages; Cancel preserves the current project, as does a read error. Unsaved changes still require confirmation before replacing the montage. Missing media are checked after the project is displayed without blocking the interface, and results for another project are ignored. The saved video destination is restored without waiting for its network folder.

Timeline and Audio: source comparisons use loaded metadata without resolving NAS paths on every redraw. Audio tracks and scene indicators use indexes. Saved waveforms, their validation and audio metadata are loaded in the background; valid waveforms are reused without restarting FFmpeg. Automatic waveform persistence into the SEP also runs in the background and skips the write if a newer save exists. Stage timings are recorded in diagnostic logs. NVENC accelerates video encoding, not SEP reading or networking; media playback still depends on NAS throughput.

Multi-GPU PCs: startup checks respect the saved NVIDIA, Intel or AMD priority. A real successful NVENC encode with the selected FFmpeg overrides an older cached QSV fallback. Changing priority invalidates previous encoder tests; QSV fallback under Automatic or NVIDIA priority is rechecked after 24 hours. In GPU, Retest GPU engines also re-evaluates the FFmpeg/FFprobe pair while processing is idle and shows the precise fallback reason. If the error reports a required NVENC API newer than the driver provides, update the NVIDIA driver compatible with your card, restart the application and retest. NVIDIA priority cannot make an incompatible driver work. The selected encoder alone does not identify the preview GPU or explain every slowdown.

Short thumbnail names: the opening card appears as start_card, the closing card as end_card and intermediate cards as timeline_card1, timeline_card2, etc., in clip order. These names are used in the timelines and overlay editor, including existing projects. Image files and their references are preserved.

Video frame rate: open this command in Timeline or Visual timeline. Choose a target rate and mode. Automatic preserves duration for a file containing an audio stream and changes speed for a file without audio; muting a track does not alter that choice. The corresponding checkbox makes this the default for new videos. Existing projects remain unchanged until explicitly applied. Apply to selection processes selected clips, or all clips if none are selected. Clear it to configure future imports only. Review the report before confirming; Ctrl+Z restores clips, settings and linked audio.

Preserve duration keeps timing, synchronization and audio at their manual speed. Fast duplicates or drops frames and cannot guarantee perfectly smooth motion. Interpolation computes intermediate SDR frames when raising the rate; it costs processing time and may introduce distortions. Lowering the rate drops frames. Missing interpolation support or HDR sources are explicitly rejected: choose Fast or cancel.

Adapt by changing speed applies the exact target/source ratio multiplied by manual speed. Examples: 25 → 29.7 = ×1.188; 25 → 30000/1001 ≈ ×1.198801; 60 → 30 = ×0.5; 29.7 → 25 ≈ ×0.841751. Duration changes. Decimal and fractional rates are accepted: 29.7 is not 29.97, and decimal 29.97 is not exactly 30000/1001. Variable, interlaced or unqualified sources are rejected in speed mode. Presentation timestamps are checked in the background.

Keep positions leaves other clips in place and rejects overlaps. Move following clips recalculates positions while retaining gaps. Changing the target rate also recalculates clips already conformed by speed without accumulating ratios; their individual mode, method and audio choices remain intact. Anchored titles follow their clip. Subtitles and independent audio retain absolute timing and must be reviewed after a duration change.

Frame rate and audio: Preserve duration applies no cadence-related audio acceleration. If you manually choose speed mode for a file with sound, linked audio follows the new duration while preserving pitch; its rhythm still changes. You may mute it for this conformation. Returning to Preserve duration restores manual speed, volume and the original enabled/muted state. Music and independent audio are not automatically stretched.

Export frame rate is independent of timeline rate. Follow timeline frame rate uses its exact value; clear this option to choose another value. A different export rate resamples frames without applying the speed ratio again. For a 25 fps montage with sound exported at 29.97: open Video frame rate, choose 30000/1001 for the timeline, Preserve duration (or Automatic for files with audio), apply to clips, then export following the timeline. Audio retains its duration. Frame boundaries are rounded on the global montage grid rather than adding one frame per clip; final audio is fitted to that global duration.

Video frame rate display and shortcuts: lists and the review show 29.97, 29.7, 59.94 and 59.4. The 29.97 and 59.94 choices use the exact values 30000/1001 and 60000/1001 respectively. For a different strictly exact decimal, enter its fraction, for example 2997/100. Ctrl+Enter starts Review before applying; Ctrl+Shift+C cancels or closes; Ctrl+K searches commands. F11 toggles fullscreen, Ctrl+M minimizes/restores and Esc leaves fullscreen or otherwise closes the window. These shortcuts also work from the frame rate fields. Applying changes still requires the usual confirmation.

Before automatic analysis: clicking Analyze and create montage first checks the frame rates of eligible videos in the camera and drone folders. If they differ, Video frame rate opens before image analysis: choose the target rate and mode, then Review and start analysis… (Ctrl+Enter) and confirm. Automatic keeps the duration of videos with an audio stream and adjusts the speed of videos without audio. Cancel stops the launch and keeps the existing montage. If all frame rates match, analysis starts directly. Metadata is read in the background and reused during analysis.

The Frame rate adaptation column, after Speed in the Timeline table, describes each clip’s processing at the timeline frame rate. Speed adjusted shows the source and target rates and effective speed factor, for example 25 → 29.97 · ×1.1988. Interpolation indicates calculated intermediate frames. With unchanged speed and preserved duration, 25 → 29.97 gives Repeated frames using Fast; 29.97 → 25 gives Dropped frames. No adaptation means no adjustment is needed. Selecting Interpolation alone does not cause that label to appear: the clip must actually need it. Manual speed and frame rate adaptation combine; multiple treatments can appear together. Still image, Unknown frame rate and Interpolation unavailable (HDR) avoid claiming an inapplicable treatment. Speed keeps showing the overall effective speed. A different export frame rate may require additional resampling. Widen the column or use the horizontal scrollbar to read long descriptions.

Checking an already applied frame rate adaptation again shows “No changes needed: these settings are already applied.” without another confirmation or undo step. The review lists only clips that actually change. Changes to the method, audio, export rate or new-import defaults still require confirmation, even if the final duration stays the same. Identical applications do not multiply the speed factor again.

Timeline GPU preview — audio startup: the audio mix is prepared directly, without building video filters or opening muted videos to produce the music. This avoids repeated, lengthy NAS media checks before the first audio data arrives. Volume, effects, fades and mixing settings still apply. If audio arrives after a silent playback interval, outdated samples are discarded before sound rejoins the current video position. Any remaining delay depends on access to the audio sources and their processing.

Private downloads: the site, assistant, catalog, components offered by the assistant, latest release and stable release require a username and password over HTTPS. An address alone grants no access. StreamerEditorPro download links are removed from public pages. In the assistant, enter credentials under Private access and click Sign in; the catalog loads after authentication. The password is not written to preferences or logs; it remains in memory for the session only. The browser and assistant each require sign-in. Stable rollback and resumed downloads use the same private access. This protection cannot revoke copies already downloaded.

Local and online AI thumbnails: the reference is decoded directly from the selected video image, without using the composed timeline render. Timeline text, logos, emojis, transitions, fades, overlapping tracks and image adjustments are excluded. Trimming, speed and playback direction determine source time; a selected scene remains the source even if another track covers it. New captures retain scene and source time. Selected legacy captures are recaptured from the video; previous AI proposals are not references. Card titles, text and logos intentionally added in the thumbnail editor remain separate layers; timeline overlays are not automatically copied into Postcard layouts. Before preview shows this clean source. Online preview makes no paid request: sending waits for Generate automatically. Local and online outputs use exact project dimensions, including UHD and portrait.

Audit of all 40 renderings: 31 photographic looks and 9 source-guided artistic effects. Photo looks use color, contrast and, where appropriate, grain, bloom or haze; they do not create mountains, deserts or new objects. Relevant palette names are clarified. The nine artistic effects combine a material guide preserving original shapes with textures from the existing local AI model, using a bounded contribution to prevent subject replacement. Comic and anime contours are lighter, pencil returns to light paper, and artificial regular painting textures are removed. Clay relief uses 12 distinct frames per second, encoded at project frame rate. Pastel tones remain intentional in soft looks only. No additional model or larger offline pack. Still preview is indicative; motion, occlusion and scaling may differ. UHD output follows project dimensions but does not mean native UHD model inference. Keyboard shortcuts and commands now uses the high-resolution StreamerEditorPro.ico icon.

Stable reference changed at the user’s request: the tested release of 8 November 2026 becomes stable. The 28 October release is removed from downloads and deleted after the new release is verified. The private catalog identifies the stable build, tests and hashes; a future release may be newer without changing this reference.

Updates: the window now uses the authenticated private catalog shared with the assistant. Enter credentials and Check; it never calls the removed public manifest. The password stays in memory only. Startup checking is deferred until a private session is open. One button offers the previous reference release, identified by version and build; the legacy version list and manual lot picker are removed from this window. Build 12 of 8 November remains the rollback reference for the next corrective release. Local and online AI thumbnails: Generate automatically creates nine proposals, three each in Beginning, End and Timeline. The first in each tab uses the chosen manual reference; the other two use different images randomly drawn across the entire montage. Sources exclude timeline overlays and results use project dimensions. Each tab has distinct proposals marked Manual selection and Random selection. An overly uniform timeline gives an explicit message before any paid request. Online generation uses nine images with one reference per image, no automatic paid retries and no generation for a still preview.

Video duration: the 4-second cap is removed. Entire scene is selected by default; enter a positive number of seconds, including decimals, for a custom duration. It is clamped to the available section after the start position, accounting for speed and reverse playback. Photo looks are streamed; artistic effects render in short bounded-memory blocks with the same seed, then assemble at project frame rate. Blocks limit memory, not final duration. Reverse video is also segmented when its buffer would be too large. Progress shows the current block and measured stage completion. Long renders take time and need temporary/output disk space. Cancel stops processing and removes temporary pieces; insertion into the timeline stays explicit.

Build 14 — Updates: all six buttons share one row. The window opens wide enough for their labels; a horizontal scrollbar keeps them accessible if it is narrowed. Escape closes the window. After successful sign-in, Windows DPAPI encrypts username and password for the current Windows account, outside the installation folder and projects. Both the update window and assistant restore them after restarting or reinstalling StreamerEditorPro under that account. An encrypted configuration backup can recover a missing file. This does not transfer secrets to another account or a fresh Windows installation. Closing the timeline hides it immediately; readers and decoders shut down in the background before their surfaces are destroyed. Display settings in standard project files save without decoding or re-encoding embedded images, off the UI thread, coalescing the latest values and protecting concurrent explicit saves. Undo history transfers without a full deep copy; Undo and Redo remain available. Resources remain protected during cleanup. Build 12 stays the stable rollback reference; build 14 is the new corrective release.

Build 15 — New videos from the local generator use the source filename followed by _AI_ and the style name in the language selected when rendering starts. Track and camera labels remain visible in the timeline. The result is always MP4; the original MP4 extension case is preserved. Other source formats produce .mp4. Existing variants are retained: subsequent versions receive (2), (3), etc. Windows-invalid characters are replaced and overly long names are shortened. The metadata file uses the same name. This applies to photographic presets, artistic AI styles and videos assembled in blocks. Existing files are not renamed. Local compilation only, without a setup, portable archive or online publication.

Build 17 — Public installation and retained licence. The website, assistant 1.4 and update window no longer request a username or password to download StreamerEditorPro or its AI components. Resumable downloads and SHA-256 verification remain. The server offers only the new application release with DEMO mode and activation; legacy unlicensed builds 13 and 14 and their publication copies are removed. At launch, this same release is the stable reference, so no older rollback is offered. All AI components remain available. Download access is separate from product activation: a valid product key is required for full features. Existing licences migrate automatically to Windows-encrypted storage without a new activation. Primary and recovery copies stay in the Windows user profile outside the program folder and survive application reinstall or uninstall/reinstall. A missing or unreadable primary can recover from a signed copy for the same machine. Revocation updates both copies. Signatures, expiry and the 30-day offline limit remain checked. This applies to the same Windows account and profile; it does not guarantee transfer after reinstalling Windows or changing computers. Do not manually delete licence data.

• An enabled-by-default checkbox in the Update window remembers the user’s choice. Manual checks remain available when disabled. An update found at startup is offered before downloading and installation, with the opportunity to save the project. Manual and history updated in all five languages.
• Rebuilt 0.20.367: HTTPS updates with mandatory SHA-256, without requiring Authenticode in the shipped configuration. Hash checked again before launch. Manually install the new Setup once; a remote manifest cannot change local policy. Local rollback packages without a digest still require signature verification.

• Compare the published SHA-256 with the Setup actually used for installation, in automatic and manual checks. The baseline is written only after installation, with messages in five languages. Install this fix once to initialize it; the channel must be published and active. Older versions remain blocked and signature checks are retained.

• Phase 1: clarified HDR/HDR10, peaks/LUFS, audible-clip priority and bilingual SRT.
• Phase 2: single auto backup, media relocation, optional local AI module and stem preservation for export.
• Phase 3: GPU/AMD scope, NAS and audio diagnostics, versions and window controls.
• Corrections within the existing 22 sections, references to documented functions and regeneration protection. No new editing feature.

• Added a 90-row reference grouped by window, with keys, their actions and usage conditions, in all five languages.

• Versions are sorted numerically in descending order, with the latest version at the top and older 0.15.xx versions at the bottom.

• Removed untitled entries from the update history. Titled versions and their release notes are preserved.

• Removed obsolete status lines from the release notes. Version numbers and other information are preserved.

• The About panel now displays the official StreamerEditorPro logo instead of a recreated mark.
• Added a five-language window summarizing the licences for FFmpeg/FFprobe, VLC/LibVLC, whisper.cpp, the optional Demucs module and Python libraries.
• The complete THIRD_PARTY_NOTICES.txt supplied with the application can also be read in this window.

• The About window now provides a detailed overview of StreamerEditorPro, its workflow and the resources included with the application.
• Features are shown as a structured list, with version and channel information grouped at the bottom of the scrollable panel.
• The manual and update history remain directly accessible; all text is available in five languages.

• The manual now includes six guided workflows: first montage, scene and timeline refinement, audio mixing, titles and subtitles, Shorts Studio, and export and recovery.
• Each workflow explains the steps, checks to make and precautions specific to the application.
• All six sections are available and searchable in the five languages.

• The User Manual, About and Update History restore their last size and screen position.
• Changes are saved automatically while moving or resizing and when the window closes.
• Maximized state is remembered separately; off-screen coordinates are brought back onto the current display.

• Added a manual section reproducing the interface explanations, grouped by window: Automatic Montage, Visual Timeline, Edit Scene, Audio, SRT Subtitles, Shorts Studio and additional tools.
• All 47 texts come directly from the application’s five-language translation catalog, matching the help shown on screen.
• The new section is searchable in the manual.

• The User Manual and Update History are no longer declared transient windows, so their title bars keep the Windows Minimize, Maximize/Restore and Close buttons.
• The F11, Esc and Ctrl+M shortcuts remain available in both windows.

• Added the F11 shortcut in the User Manual and Update History, without adding a button in their interface.
• Escape exits full screen and restores previous window state; out of full screen, it closes the window.
• Ctrl+M retains the enlarged/restored alternation with the native title bar commands. Explanations are available in all five languages.

• The manual now uses the native Maximize/Restore buttons in its title bar; Custom magnification buttons and F11 full screen have been removed.
• Ctrl+M toggles between the maximized state and the previous window, as shown in all five languages.

• Added an About window showing the graphical presentation of StreamerEditorPro and displaying the identity, version and main functions of the application.
• Integration of the complete update history, viewable and searchable in all five languages, from all changelog entries.
• History is accessible from Tools and from About; the contents remain included in the standalone distribution.
• Tests check the completeness, languages, search and consistency of the versions displayed.

• The manual can go to full screen with F11 and return to windowed mode with Esc; the window buttons also offer full screen and enlarged/restored window alternation.
• Ctrl+M toggles between the maximized manual window and its previous size. The shortcut and these modes are explained in all five languages.
• Tests check display transitions and window state restoration.

• Added a detailed manual in French, English, Italian, Spanish and German, with search, navigation by sections and graphic diagram of the StreamerEditorPro interface.
• The manual is accessible by Tools → User manual or F1 and follows the active language.
• Guide files are included in the standalone compilation.

• The startup logo now follows the same size rule as StreamerEncoderPro: 32% of the smallest side of the screen, with a limit of 280 to 420 pixels.
• Transparent margins are cropped before resizing, as in the reference application.
• A test verifies the rule on several resolutions.

• Projects now keep a single backup copy with the suffix .auto.sep, replaced with each successful backup.
• Old chains of nested .previous copies and .auto duplicates are cleaned up after publication, retaining the current project and its unique backup.
• Tests cover multiple successive backups, restoring the previous version and cleaning up nested names.

• The “Show/Hide settings” handler accepts the Tk event transmitted by Ctrl+R and Ctrl+Shift+R without argument errors.
• The shortcut hides or redisplays the settings and consumes the event to prevent it from being processed a second time.
• A test reproduces the call with a keyboard event and verifies the masking of the panel.

• The test protocol distinguishes two first NAS tests before full reading of the media of the randomized matrix after local copy; VLC cache size is indicated in each campaign.
• The final executable retains the Auto-to-CPU one-time fallback, the seek and the read/pause intention; the overall Auto preference remains unchanged.

• The placement matrix logs the executable version and footprint, GPU adapters/drivers, VLC, media characteristics and footprint without exporting its path.
• If Auto fails after a recovery attempt, only the player of this scene is reopened once in CPU at the last requested point; the overall GPU preference remains unchanged and the fallback state is visible in all five languages.
• Start, pause, seek, resume, recovery and end of media operations are serialized outside the interface thread with deadline and rejection of late returns.
• Actual VLC decoder and exact visual timestamp remain unconfirmed; Auto/NAS compatibility is not announced as resolved.

• Inspecting a scene corrects a position overrun up to twice by pausing VLC and then re-requesting the exact timestamp; a distant image is never accepted as a target.
• Resumption keeps the last positioning request, even if it is more recent than the position confirmed before the incident.
• The reader is parked as soon as it is opened in the background, before being returned to the interface.
• Diagnostics distinguishes read requested, inspection, VLC clock, delivery/adoption delays, UI loopholes and slow native calls.
• VLC restarts and changes between source/proxy are serialized per player and executed in the background; closing no longer holds the window waiting for native release.

• The Audio tab releases its shortcuts and transport bindings when closed, stops its polling, and undoes deferred drawings and actions before Tcl is destroyed.
• Waveform redraws use stable Tcl command; any remaining idle task that reference is canceled before this command is released.
• Audio UI tests destroy windows and free their references before collecting on the owning Tk thread.
• The isolated launcher imposes a delay per module, stops the process tree in the event of a blockage and distinguishes native crash, absent unittest summary, residual Tk callbacks and counted skips.

• Added an encoder register that separates provider and codec, with AMF SDR H.264 and HEVC profiles using options specific to the supplied FFmpeg.
• Auto-selection attempts AMF only after actual probe detection and encoding; an imposed AMD profile clearly fails if the driver or encoder does not work, without switching providers or codecs.
• Added AMD Radeon preference, AMF boot status and explicit AMF choices in five languages. Batch output validation correctly recognizes H.264 and H.265/HEVC.
• HDR AMF, AV1, proxies and Shorts remain outside this first perimeter. The presence of a Radeon and real AMF encoding remain to be qualified on an AMD computer.

• The startup check only displays the validated encoding engine: NVIDIA NVENC, Intel Quick Sync or CPU, without mentioning the engines tested and then discarded.
• Results kept in the old cache are presented with the same simplified wording.
• Removed the alert that detailed the NVENC failure when QSV or CPU was held.

• Ctrl+Shift+O opens or hides the advanced “Edit Scene” options.
• The shortcut is recognized and displayed in the button's help depending on the language: French, English, Italian, Spanish or German.

• Physical clicks received by the elevator but not identified by ttk now use a fallback limited to unrecognized elements: arrows with repetition stopped on release, page scrolling and cursor sliding.
• Interactions that ttk correctly identifies remain managed by native bindings.
• Added a regression test that forces blank identification and validates clicks, holds, releases and drags.

• The recalculation of the scrollable area is also triggered when the canvas changes size (maximization, restoration or movement of the separator), even if the settings have not changed.
• The interaction remains entirely native ttk, identical to the visual timeline.
• Added a test of enlarging then reducing the panel in the clam theme.

• Removed the track click handler specific to “Edit Scene”, which competed with the mouse bindings of the ttk theme under Windows.
• The bar now only uses native ttk interactions, such as that of the visual timeline; the explicit calculation of the scrollable area remains active.
• Tests of arrow/track clicks, sliding and successive reopening of windows.

• Direct clicking on the track now distinguishes the physical areas of the arrows even when the Windows theme incorrectly identifies them as the track.
• Arrow events remain transmitted to ttk; clicks in the track retain direct movement.
• Test simulating a theme which returns “trough” on the arrow pixels, with verification on two successive windows.

• A click in the track no longer blocks the ttk event when the window is still finishing its scrollable area calculation; a new calculation is requested immediately.
• Deferred elevator recalculations are canceled when the panel is closed to avoid obsolete callbacks between two “Modify scene” windows.
• Click test up and down on two panels opened successively.

• Left-clicking in the “Edit Scene” slider track now places the cursor directly at the clicked position, both up and down.
• Arrows and cursor drag remain managed by ttk, as in the visual timeline.
• Reinforced mouse test with direct jumps down then up.

• The “Edit Scene” panel calculates its scrollable area from the actual height required by its settings, even when controls are added after the preview is created.
• A deferred recalculation after layout allows you to use the scrollbar without having to open “Advanced options”.
• Test added for controls added late; Elevator panel tests passed.

• The displayed size of the startup logo is reduced by 20%, maintaining its centering and proportions.
• The original image file remains unchanged; Resizing is applied only to the Start screen.
• Splash dimensions test adjusted to validate the new scale.

• The vertical bar in the “Edit Scene” settings now delegates clicks, arrows, track clicks and drags to native ttk interactions, such as the visual timeline scrollbar.
• Removed local mouse handlers that prevented native arrow visual states and actions from working correctly.
• Tk tests covering up/down clicks, track click, drag and wheel.

• The “Next” and “Previous” buttons now retain the preview interface and reuse the already open VLC player when changing files.
• The metadata and smooth preview of the new source are relaunched in the background; the results of an old analysis are ignored after navigation.
• Added tests checking preview and player reuse during scene changes.
• Successful targeted testing and Python compilation. The complete suite remains limited in this environment by missing FFmpeg and 4 failed asset lifecycle tests.

• Left clicking in the elevator track immediately repositions the cursor and associated settings; the arrows move by one unit.
• Left cursor dragging is now handled explicitly based on the actual dimensions of the rail and cursor, including in the vertical separator of “Edit Scene”.
• The wheel keeps its current behavior. The fix is ​​validated by Tk mouse events on the real PanedWindow structure of the scene.
• Clean-Windows blockers remain listed in the qualification register.

• A click in the elevator track directly repositions the settings; Cursor dragging is handled explicitly to remain reliable with the Windows theme. The wheel remains available.
• Ctrl+R shows or hides scene settings in five languages; Video reversal now uses Ctrl+Shift+R.
• The Timeline progress bar is placed to the left of the save information to better preserve their width.
• Verification of clean-Windows blockers kept in the project risk register.
• Successful targeted testing of elevator, shortcuts and translations; Full validation required before packaging.

• The bottom panel of “Edit Scene” now has a vertical scrollbar on the right. It allows you to reach advanced settings and options when the window is too low.
• The mouse wheel scrolls this panel, while the preview area maintains its own navigation.
• Zoom on the visual timeline wheel is accelerated by approximately two times per standard stop, while maintaining proportional adjustment on the precision wheels and the anchor point under the mouse.
• Added an interface test checking the bar, scrollable area and wheel.

• Audio waveform vertices now keep their X/Y pairs when flipped to close the polygon. This removed the huge orange/gray triangle that covered the timeline.
• Edit Scene settings and navigation controls are placed in the bottom panel of the splitter, instead of using the video canvas as a pack neighbor while it is handled by a PanedWindow. Opening a scene no longer triggers the “canvas isn’t packed” error.
• Added regression tests on waveform polygon boundaries.

• The timeline now imports the audio peak color function used to display estimated levels; the absence of import triggered an exception on each redraw.
• Identical Tk errors repeated within a five second window are logged without opening a new modal box with each callback.
• Targeted validation: drawing regression and dialog limitation tests, plus relevant timeline tests. No portable rebuild/setup without “go”.

• The video preview and the control area are separated by a movable handle; The minimum height of the video adapts to the Windows display scale.
• Scene settings and advanced options are collapsible and their state is memorized. Diagnostics and cache/decode settings are hidden at startup.
• The separation ratio is remembered as a proportion to adapt to changes in screen size and scaling.
• Full screen video retains its behavior by temporarily hiding controls, then restores them on exit.
• Targeted verifications in progress; no complete compilation or portable/setup generation before the explicit “go”.

• Audio level analysis now reports its progress during FFmpeg decoding and remains overrideable, with separate duration readings for decoding, audio and pass selection.
• Auto-editing steps (colors, pacing, subtitles, construction, normalization and thumbnails) display their progress and durations in the log.
• Thumbnails from the same video are extracted in groups with a single PyAV player and kept in a limited local cache; the FFmpeg fallback remains available for incompatible formats.
• Full previews with effects are cached by media, revisions, project and rendering profile, with atomic writes and storage limit.
• Targeted validation: 126 tests pass. No complete compilation, portable application or setup has been generated, in accordance with the instructions to wait for the “go”.

• The Timeline window is displayed before its first drawing; the video thumbnails are loaded progressively and kept in a local cache limited to 512 MB, invalidated when the source file changes.
• An isolated update of a sequence only repaints its block if its position, tracks and the rest of the layout have not changed. Structural changes continue to update dependent elements together.
• The arrival of a waveform or thumbnail only updates the element concerned, without rebuilding the entire timeline.
• Confirmed encoding engines are stored by computer and version of FFmpeg; a translated button allows you to restart the tests. The analysis avoids retrying a QSV engine that has already exceeded its deadline, until this new test.
• Targeted tests pass; the complete compilation, the laptop and the setup remain waiting for the explicit “go”.

• The waveform of each audio track is rendered with a compact polygon capped at 800 points, instead of thousands of graphic objects recreated with each reconstruction.
• Opening the VLC player in “Edit scene” goes into the background; The window remains responsive during initialization and access to network media.
• Diagnostics now separates drive creation and open times, and details timestamps, frame counters and states when a position expires.
• In continuous playback, a new image actually displayed confirms the positioning even if the VLC clock has already exceeded the estimated window; Paused inspection maintains precise control of the requested time.
• Export blocks can cross scene changes and their duration adapts more to the available memory, reducing FFmpeg restarts of edits of many scenes.
• On Intel machines where the hardware test passes, Quick Sync also processes contrast, brightness, saturation, sharpness and scaling in VPP before transferring to system memory; the CPU hybrid path remains enabled if this filter is refused.
• Export logs now measure the preparation and actual duration of each block separately.

• After 750 ms without audio samples, the silent video clock takes over. A later returned audio stream is realigned to the position played.
• Video playback continues if audio output is unavailable; The exact error is displayed and persisted in the diagnostic log, along with further connection attempts.
• Pause/Resume restarts buffering without losing position. The validation compilation does not generate the laptop or the installer.

• Stable release remains blocked by unsigned artifacts, independent clean-Windows qualification, and any exact-candidate campaign that cannot be completed on this host.

• Revalidated generated video-selection corpus and HDR/export tests while preserving the tolerant selection profile and the distinction between HDR export checks and physical-display qualification.
• Real annotated camera/drone media, multi-vendor hardware, and visual review remain explicitly unqualified on this host.

• Extracted backend-independent ordered native shutdown/retry into a small lifecycle contract; the player remains responsible for deadlines and thread supervision.
• This is a bounded extraction; command/undo, timeline transport, and the wider UI remain incremental architecture work.

• Updated the current risk register from M00–M10 with implemented evidence, unresolved exact-candidate/clean-Windows blockers, and accepted scope limits.
• N1 (0.20.310): first SharedMemory allocation failure now makes live product processes visible as unknown to external collection until publication recovers.

• VLC reader shutdown now retries only the unfinished stop/release stages after an error, while preserving bounded waits and supervisor reporting.
• Added fault-injection coverage for retry after instance-release failure without releasing the player twice.

• Package-lifecycle evidence explicitly marks the configured development host and standard-user/clean-machine qualification as unverified.
• Added a regression check preventing a development-host run from claiming independent Windows qualification.

• Added a focused release-gate test proving that absent, historical-hash, and failed campaign reports cannot qualify the current candidate.
• Exact hardware/endurance campaigns remain a separate final-candidate activity and are not inferred from these local checks.

• Playback frame and dropped-frame totals are reconciled against categorized events; startup latency aggregates are checked against every expected generation.
• The concurrency qualification recipe now records per-frame playback/startup/seek scope, dropped-frame partitions, and first software-audio data for every generation.

• The loss ratio only uses new frames during continuous playback; boot and seek images are counted separately.
• Recompilation and final validation restarted before replacing the Laptop and Setup.

• Adaptation of the historical integration test to the extracted snapshot service; assertions now check the contract and supplied domains.

• Decoded annotated corpus reexecuted; strict selection and calibrated confidence remain deactivated due to lack of sufficient real corpus.

• Verified road matrix: HDR export available, native/physical HDR display not claimed and explicitly tolerant video selection; re-controlled language coverage.

• Extraction of the owners' snapshot contract into a reusable service; The UI orchestrator marks pending before calling it.

• A discovery interrupt, an inaccessible slot or an incomplete slot closure now produces a blocking unknown state; multi-owner matrix added.

• Measurements of first audio data, first image and startup silence separated by generation; three bounded budgets added to the profile.

• The gate keeps the refusal stable as long as a blank Windows VM in standard account has not been executed; local compilation and smoke are not presented as this external qualification.

• Targeted correction of UTF-8 encoding of capabilities profile limits; contract locked by test.

• Loss ratio now compares lost frames to new and lost content frames; repeated swaps cannot dilute it anymore.

• The shared state becomes pending before any copy of the project; MemoryError, snapshot failure and retry are covered.

• Protection publication invalidated before serialization: a capacity refusal also blocks an external collector. No disk access to interpret the indicated volume.
• Historical recipes linked to their exact AI archive; Updated scene playback fixture for supervisor.

• Correction of the encoding of the risk register before qualification and delivery of the exact candidate.

• Final validation awarded to the exact candidate; real CPU/CUDA campaigns, endurance and deliverables. Optional module recipe with installation, repair and removal in an isolated profile.

• Annotated corpus of generated video files: movement, stopping and cutting tested by real decoding. Tolerant profile maintained, no strict trust invented.

• Reference independent of the stereo mix (gain and mute track) compared to the exported PCM.
• File protection diagnostic actionable and tested in five languages. Hardware HDR routes and human language review remain separate.

• Supervised waveforms and scene polls; sources/snapshots retained; cancellation on change of project.
• VLC playback: retained files and bounded native closure with diagnosable thread. Inventory distinguishes explicit supervisors.

• Risk register: current reference, G00/G07 identifiers reconciled with explicit mapping, updated competition reasons. Historical evidence preserved.

• Strictly distinct presentations, image numbers and generations; centralized thresholds.
• Software silence during playback, startup, seek cuts and deviation of audio/video clocks measured separately from hardware underruns.

• Windows standalone recipe: native capture of diagnostics without Python dependency and 180 second deadline. Development position explicitly not qualified as Windows virgin.

• Proposed module: fully verified public v3; worker 0.20.278 and real phases. v2 remains an authorized local variant.

• Competition recipe: two-minute AI source and one-minute preview; Longer actual active window, without prolonging engine events.

• Publication health in system memory: an error session blocks collection, even if the disk refuses a lease. Explicit retry and closure.

• Corrected competition instrument: non-blocking observation, separate continuous reading of seeks, cumulative losses and catch-up losses retained.
• Explicit seek budget at 1500 ms, p95 reading at 100 ms, reading losses limited to 5%.
• CPU/OpenGL/physical output route declared; new identity of the candidate after correction of the campaign.

• Reinforced final gate: configured budgets, ordered AI/PID phases and presentation samples in the inference window.
• Windows own standard user required for distribution qualification.
• Register G01–G12: data counterexamples and closure corrected; competition, Windows own and signature remain blocking.

• Tolerant profile displayed in the five languages; heuristic score distinct from calibrated confidence.
• Source terminals validated for camera and drone candidates; cache invalidated for this new contract.
• Stable/cut/freeze corpus and unfinished cases; the strict precision mode remains explicitly unqualified.

• Independent BT.2020 NCL references for HDR corrections in PQ code space, maximum error less than 3 10-bit codes.
• Real stereo audio mix: full PCM and exact duration at 48 kHz.
• Replayed CPU/GPU transition and karaoke audio regressions; physical HDR display and full hardware matrix not qualified by these tests.

• GPU readers explicitly attached to the application supervisor.
• Closure with shared deadline, inventory of direct and descendant process threads.
• Common transport terminals for audio and video; refresh of owners after history synchronization.

• Completed catalog for 43 messages and titles in FR/EN/IT/ES/DE.
• Preflight HDR announces unqualified transitions, wraps and fades before rendering.
• Multi-line messages translated line by line; Inventory classifies user messages and monitors their coverage.

• Standalone video/audio import, save/resume, CPU export and preview decoding, without Python system.
• Own Windows validation distinct from the local smoke test; required external environment.

• AI phases issued around actual operations with PID, order and timestamp.
• Visible SDL qualification, accumulation of losses and measured competition windows.
• Blocking budgets: images in inference, gestures, presentation and audio.
• The modified worker requires a new archive; remote publication to requalify.

• A release error preserves the operation result and its terminal notification, with separate cleanup diagnosis.
• Closing an owner can be retried after failure and shares an explicit deadline.
• Processes registered during or after the shutdown are immediately stopped and remain inventoried; expectations share the overall budget.

• Recovery, purging and counting deny links and reanalysis points in each component and in generations.
• Paths are double-checked under the lock before reading, moving and deleting.
• Non-root controls remain intact in junction tests.

• Each visual history publishes a separate domain, preserved across major refreshes.
• Closing transfers the histories then frees only the domain concerned.
• Behavioral tests check refresh, collect, two windows and release last owner.

• Managed tasks publish their assets under an owner linked to their identity and automatically release this owner upon completion.
• The critical processes of analysis, AI separation, audio probing and export are attached to the current task, then completed and unsubscribed when closing.
• A result arriving after the start of the closure is rejected even if its project and its revision still match; the reason remains in the diagnosis.
• The shutdown report distinguishes residual tasks, processes and threads. The architectural inventory identifies now-managed processes and direct owners that still require bounded closure.

• The HDR matrix now separately controls luminance, both chroma planes, neutral path and corrections in 16-bit RGB space.
• HDR10 video transitions, overlays and fades remain refused until they are qualified; the user receives a clear explanation in five languages ​​before execution.
• Audio, Shorts, Layouts, and Retrieval user recipes include new language controls and a reproducible inventory of non-catalog direct messages.
• Milestones 0.20.245 to 0.20.261 are placed in history based on tests and proofs already stored.

• The quota of a parent folder lists all the authorized quarantines of child projects, without counting twice a quarantine copied to another.
• Data, logs and temporary are counted separately; a route error or a junction point blocks any destructive collection.
• Recovery and purge cycle through child quarantines under the same transaction lock and maintain containment at the authorized root.

• Qualification distinguishes worker startup, loaded model, and active AI computation, then measures each preview priority window.
• Qualification playback feeds a real audio device, counts callbacks and underruns, and compares multiple actually presented frames instead of a single non-empty framebuffer.
• The governor targets the selected GPU, exposes its index, UUID and name, counts its VRAM and suspends/resumes the entire AI process family.
• The report explicitly describes the qualified policy: priority to preview by suspending the AI.

• The concurrent proof is only admissible if the hash of its worker corresponds exactly to the executable extracted from the configured IA archive.
• A common WAV validator reads the entire PCM in bounded blocks, rejects truncated files and produces a fingerprint of the actually decoded stream.
• The four CPU/CUDA stems must exactly match the frequency, channels and number of samples of the source; the declared tolerance is zero samples.

• A durable and updatable owner protects the assets of the living project, Undo/Redo stacks, clipboard, drives and pending batches.
• Each lease records the birth identity of the process and the operation, so that a reused PID is not mistaken for the original owner.
• The timeline republishes its inventory after each snapshot, restore, history transfer and closure. An asset becomes collectible again only after the last owner.

• The final candidate brings together attributable evidence of user journeys, 10-bit HDR accuracy, endurance, and real-world concurrency between OpenGL preview and HTDemucs CPU/CUDA separation.
• The Setup cycle takes the normal user path with explicit acceptance of the license, runs the self-test of the installed application, uninstalls the product then checks for the absence of residue.

• Shorts and Audio proofs now replay real images and all produced PCM; a simple file presence or container poll is no longer enough.
• The automated recipe brings together the Shorts, Audio, batches, migrations, recovery, shortcuts, layouts and text courses for the five languages.
• The HDR test pattern compares the decoded levels of the source, the real graph produced before encoding and the final HDR export. The gate requires this proof of accuracy in addition to the tags.
• WAV validation reads the PCM in bounded blocks. The Setup cycle will use the normal user path with explicit acceptance of the silent license, without the old /QUALIFICATION switch.

• The concurrent campaign now launches the strictly compatible HTDemucs 0.20.243 worker, executes a CPU separation then a CUDA separation and validates the four stems produced.
• During CUDA, the actual preview decodes media, feeds audio callbacks, renders images via the OpenGL composer, performs five seeks, and manipulates a timeline of 1,000 clips.
• The governor effectively suspends the CUDA worker when the preview starts and then resumes it after releasing the interactive priority. These events and process family resources are recorded.
• The proof carries the hash of the worker and that of the exact public AI archive; a simple self-test can no longer satisfy this level of qualification.

• The capabilities profile now has a strict schema: an empty, incomplete or poorly typed document becomes a qualification block and the application returns to a safe SDR profile.
• Each gate requirement indicates validated, not-required or failed and remains assigned to the exact hash of the sources, scenario and deliverable concerned.
• A worker self-test no longer constitutes proof of inference. The concurrent campaign must employ the common handshake validator, the exact compatible worker and declare the inference and its output separately.
• HDR accuracy, concurrency level and actual installation mode are now explicit requirements of the delivered profile.

• Nested projects share a canonical transaction lock per volume. Registrations and deletions of leases are now part of the same transaction as collection.
• Persisted leases are always merged with the inventory provided by the caller; an empty local list can no longer hide an active owner from another instance.
• Quarantine writes and synchronizes its log before any movement, restores immediately after an error and knows how to recover the prepared or moved states.
• The quota distinguishes between active generations, quarantine, temporary ones, bytes moved and the space actually freed. The definitive purge becomes a separate operation with a retention period.

• Neutral HDR rendering now entirely avoids the eq filter which negotiated an 8-bit YUV intermediate. A ramp of 1,024 levels therefore remains a ramp of 1,024 levels before encoding.
• Contrast, brightness, saturation, sharpness and HDR corrections use a 16-bit RGB workspace with explicit PQ/BT.2020 conversions, then revert to limited 10-bit YUV420.
• Neutral processing is removed from the graph and HDR evidence now distinguishes signal metadata from actual pixel precision.

• Candidate identity includes capability profile, D00–D12 register, actual packaged Python dependencies, runtime requirements, and future native binaries without circularly integrating output reports.
• The final lock approximates the complete suite, user journeys, interactive concurrency, endurance, HDR10 exported by the Renderer, packet cycle, exact AI module, public body and signatures according to the announced capabilities.
• The Setup has a silent mode reserved for qualification and allows an isolated user test; the exact package is installed, self-tested and then uninstalled without retaining a false success.

• The Shorts desktop and external process scenarios are executed with the distributed binaries. The NVENC proxy now uses a compatible short GOP and no longer silently falls back to the source after an initialization error.
• An assignable matrix replays Audio Studio paths, stems, racks, solo/mute, links, Undo/Redo, preview/export, batches, migrations and five languages.
• Shorts and decodable audio outputs are retained with size and SHA-256. Batch errors, unavailable destinations, cancellations, and changed entries never produce a false completed status.
• The own machine recipe is versioned and deferred to the exact package of phase F, since the Portable and the Setup must only be created in the last phase.

• A campaign attributable to the candidate simultaneously exercises decoding and seeks, timeline gestures with 1000 clips, audio exports, interactive priority and self-test of the installed AI worker.
• The reports distinguish the requested sampling interval from the cadence actually observed.
• VRAM scope is now named: NVIDIA Compute Process Accounting, which may exclude some presentation allocations.
• The campaign also checks the release of the preview priority and the absence of late exit applied after the tested context.

• The package explicitly declares its HDR capabilities. Direct3D HDR10 preview is hidden when its native DLL is not compiled; HDR10 export remains available.
• A future declaration of native HDR preview makes the DLL mandatory at package build and self-test time.
• The qualification target uses setparams to actually preserve the BT.2020 primaries and the PQ transfer, then an export of the real Renderer is probed in turn.
• HDR report maintains order, FFmpeg/FFprobe versions, encoding log, replayed metadata and physical validation limits.

• Projects, histories, previews, exports and workers can publish their references in interprocess asset leases separate from those of the AI module.
• Saving the project, publishing a build, and collecting use the same interprocess lock. A generation that has become referenced can no longer be deleted between the inventory and the backup.
• Cleanup discovers projects in subfolders, checks paths and junctions, provides an encrypted preview, and moves orphans to a recoverable quarantine after explicit confirmation.
• An interrupted shutdown can restore quarantined generations to their original location.

• Authenticode verification now uses a single policy in application and post control. A configured list of fingerprints becomes mandatory and can no longer be bypassed by the name of the publisher alone.
• Generation cleanup remains blocked when the inventory of living owners has not been explicitly provided, even if no saved project references the cache.
• HDR qualification accepts different versions of PyAV without masking the initial decoding error and retains this cause in its report.
• The qualification register now follows findings D00 to D12 of audit 0.20.255.

• The uses of the IA module publish the identity of the parent process and the worker; an unknown identity retains the lease and blocks any destructive maintenance. Closure collects all decisions before stopping tasks.
• The CUDA qualification reads the entire PCM body. Signatures use an exact publisher identity and pinned imprints take priority. The release campaign requires the complete and non-empty inventory of tests.
• Candidate identity covers sources, configurations, dependencies, recipes, resources and distributed folder. HDR levels, decode/conversion times, and resource peaks are published separately, with unexecuted capabilities reported as such.
• Generations of stems have an inventory, an informative quota and explicit cleaning maintaining current/previous projects, shared references and leases. Stem migrations and errors are available in all five languages.

• The collection keeps the current, previous, shared or rented generations and remains blocked when the references are not certain.
• Project and stem migration errors are specific and available in all five languages; Direct owners of threads and processes appear in the inventory.

• Actual charges separate image recovery, conversion and pipeline, as well as different HDR qualification levels.
• Endurance samples resources during operations instead of measuring only before and after.

• CUDA validation reads the entire PCM, certificate fingerprints dominate publisher names and partial suites can no longer qualify a broadcast.
• Candidate identity includes construct and trust root entries; authorized skips remain visible.

• A living owner or owner of unknown status retains his lease; only an owner shown to be dead can be automatically cleaned.
• Aborting a shutdown no longer prematurely destroys the task manager.

• The smoke test produces real media and the autonomous self-test confirms the identity 0.20.250 of the candidate.
• The endurance campaign and public verification of the AI ​​module are kept with their limitations, without presenting them as a physical HDR screen qualification.

• GPU probes are bounded and shared between snapshots. The qualification load uses real media and explicitly publishes the physical HDR validation limit.

• Any copy of stems is revalidated by its exact identity before publication; old caches require explicit re-separation.
• Two saves keep their generations immutable and a JSON failure leaves the last project open.

• Truncated logs can no longer qualify a sequence; CUDA proofs require matching audio outputs and separate durations.
• Diagnostic secrets are redacted and the signature must belong to an authorized publisher.

• Telemetry errors no longer hold the AI lock; interprocess leases exclude concurrent maintenance and cache waits remain cancelable.
• Deferred results are revalidated against the project at the time of their application.

• WAV and video exports use the requested stems, refuse missing generation and only use the original source after explicit bypass.
• The history shares a budget between Undo and Redo and reserves the target before any eviction.

• The complete isolated suite, audit of the five languages, standalone compilation and self-testing of the distributed file are executed before the creation of the final deliverables.
• The endurance campaign records the memory tray of the entire process family, the p95/p99 cost of gestures and the Undo/Redo budget. Its real duration and limits are preserved in the proof, without presenting it as a 30-minute material campaign.
• The Laptop and Setup 0.20.244 are produced only after the last validation, in accordance with the requested procedure. The optional AI module remains a separate 0.20.243 compatible package, verified by the 0.20.244 application.

• The optional module was rebuilt with Demucs 4.1.0, PyTorch 2.14.1+cu130, CUDA 13.0 and verified offline HTDemucs model. Its archive of 2,114,084,461 bytes is linked to the configuration by SHA-256.
• The qualification extracts the exact ZIP into a new folder, executes its worker 0.20.243 and links the CUDA proof to the fingerprints of the archive, the worker and the four stems. A real CUDA split of 30 seconds source was performed on the RTX 4060.
• The lock no longer confuses the build folder with the distributed package. The mere presence of WAV is no longer enough: version, duration, CUDA device, sizes and full footprints are checked.
• Remote control now requires the binary identity of the ZIP, in addition to the type and size. The server still exposes the old packet; the stable channel therefore remains closed without silently replacing this proof.
• Diagnostics recursively hides API keys, passwords, secrets, permissions, cookies, tokens and URL parameters in structured objects as well as in free texts.

• The state of the IA module is defined by a pure service separate from Tk: absent, operation in progress, model missing, repair required, CPU ready, CUDA ready, preview priority, valid, stale, missing or invalid tracks.
• The Install, Repair, Delete, Separate and Mix buttons are only active when their action is actually possible. The interface no longer displays a false ready state when the model or CPU qualification is missing.
• Structured worker errors are converted into user instructions for CUDA memory, driver, model, media, and preview priority, without directly exposing the technical output of the engine.
• New dynamic states and error messages are provided and monitored in French, English, Italian, Spanish and German. The catalog now has 1,558 complete identifiers.

• The governor measures the memory, threads and handles of the entire process family as well as the CUDA VRAM allocated to the application. The preview keeps an explicit reserve of RAM and VRAM.
• A background CUDA split now frees its worker when the active preview runs out of VRAM, then can return to CPU with a structured pattern. A simple suspend no longer retains a GPU allocation that compromises playback.
• The Undo and Redo stacks share an overall budget of 64 MiB. Rebuildable caches are excluded from the calculation, isolated entries that are too large are rejected and the measurement no longer serializes states with pickle.
• The arrival of an already visible thumbnail updates its canvas object without rebuilding the timeline. The playhead remains independent and diagnostics releases p95, p99, targeted updates and full rebuilds.
• A reproducible bench records machine profile, GPU/FFmpeg versions, process family memory, cost of 100 gestures and declared criteria. The short phase proof is explicitly marked as baseline smoke, without claiming a hardware endurance of 30 minutes.

• Export recovery and batches now identify each media, stem and FFmpeg binary with a full SHA-256. A modification outside the old sampled areas invalidates the recovery even if the size and date are preserved.
• Version 3 checkpoints link the full snapshot, strict sources, engine, encoding profile, colorimetry, precision, audio graph and a secure temporal boundary. Their manifesto is written atomically.
• Each completed batch release has its own integral identity; a file replaced after success reverts to failure. The pending, in progress, successful, interrupted, and failed states remain persisted independently.
• The final publication remains atomic and occurs after a survey. Validation requires codec, resolution, frame rate, duration, 48 kHz audio stream and, in HDR10, 10-bit pixels, PQ transfer, BT.2020 primaries and non-constant BT.2020 matrix.
• Tests cover out-of-sample mutations, checkpoints, batch restarts, late collisions, undo, stem PCM, continuous fades, segmented pixels and real HDR10 signaling.

• Each stem now has an immutable manifest with role, SHA-256, size, number of samples, frequency and channels. Portable collection, moved reopen, cache reuse, and strict exports verify this complete identity.
• Cache generations are serialized between processes by key. Postback keeps the old generation until commit and explicitly restores it if the replacement fails.
• Asynchronous results use a common envelope containing task ID, project ID, initial revision, revision policy, result, error and cancellation. The interface rejects any event coming from a replaced project, an old task or a revision that has become incompatible.
• Coordinated shutdown prohibits new jobs, cancels tasks, terminates child processes, waits up to five seconds, and logs owners that are still active. The current autosave is waited for and the final cache cleanup completes before the process ends.
• Testing covers replaced but decodable WAVs, portable builds, stale events, in-place tasks allowed, and shutdown with closed admission.

• The worker's output reader immediately renders short messages and keeps a bounded log. Breaks imposed by the governor are subtracted from the periods of inactivity and active duration.
• The worker publishes a startup, heartbeats and structured errors. CUDA to CPU fallback now depends on explicit CUDA code; a media or model error no longer automatically triggers a lengthy second processing.
• The qualification strictly checks protocol version, worker version, CPU capabilities, presence of the HTDemucs model, revision and SHA-256. The status distinguishes between worker present, CPU usable, CUDA available, model verified and separation ready.
• Consultation of the status has no side effects. Each operation carries a UUID, a PID and its timestamps; Explicit recovery on startup keeps any operation alive and only cleans up the staging of an identified dead owner.
• Tests cover small exits, silence, long pauses, incompatible handshakes, no pattern, live operations, and selective crash recovery.

• Dedicated Instrumental and Acapella exports deactivate bypass only in their working copy. The project retains its listening state and the rendered PCM file contains the requested stems, without silent return to the original source.
• AI cache keys are limited to an opaque name compatible with legacy projects. Any folder traversal, absolute root, drive letter, symbolic link, or junction going out of cache is denied before the first write.
• When a decoding queue does not stop within the allotted time, the native container remains under the responsibility of a deferred collector. It is no longer closed or repositioned while the decoding thread can still use it.
• Phase tests cover bypass export PCM content, hostile and compatible cache keys, as well as races between close, seek and blocked decode.

• The final qualification addresses the lightweight application and the optional AI module separately. The report attributes each result to the source hash, tests, executable, Portable, Setup, and, where it exists, the exact AI archive.
• A broadcast lock controls the final address, type and size of the AI ​​ZIP, its fingerprint, Authenticode signatures and channel. A stable version can no longer be released if any of these checks fail.
• The Portable and the Setup are only generated after the complete suite, the audit of the five languages, the preparation of the manifests, the PyInstaller compilation and the autonomous self-test.
• Optional module v2 is rebuilt separately with Demucs 4.1.0, PyTorch 2.14.1+cu130, CUDA 13.0 and verified HTDemucs model. Its worker 0.20.236 fixes missing console streams and produced four valid stems during an actual CUDA split.
• The Visual Timeline now accepts the mouse wheel on the timeline and on the Zoom slider. The adjustment is progressive, remains limited and maintains the same moment of editing under the pointer.

• Tasks now share unique IDs and project review. Shutdown actually waits for threads within the expected limit and then terminates the remaining processes; the previous premature return has been deleted.
• Joint project, transport and export contracts introduce stable actions independent of the wording displayed. Karaoke modes and Shorts subtitle choices use industry identifiers, with presentation translated into all five languages.
• Shorts Studio consults the shared resource governor. Its background conversions pause during an interactive preview, while the decoding requested by the scrub remains prioritized and reports this activity to the rest of the application.
• Diagnostic archives centrally redact home folders, API keys, tokens, passwords, Bearer headers and URL parameters, including in copied logs.
• The visual qualification matrix now covers 100, 125, 150 and 200%, small windows and UHD. User texts remain literal and language changes do not modify business values.
• The README describes the lightweight product, the optional AI module, the portability of the cache and the intermediate compilation and then final packaging process.

• Valid stems are collected in the project's neighboring cache when saving. The project file maintains relative references and automatically finds tracks after moving or renaming the project.
• The identity of an export resume now includes the contents of the stems, media and FFmpeg binary, as well as the complete project, the AI ​​engine, its model, its options and the output profile. An incompatible recovery is automatically excluded.
• Batches remember the identity of the media and stems during their preparation. A replaced source or modified AI track interrupts the affected project instead of silently reusing the old state.
• The duration of the export blocks adapts to the definition, the number of video tracks and the available memory. The required workspace is estimated before the first block and checked again during assembly.
• The final file is checked before publication: codec, definition, frame rate, duration, audio presence and HDR10 signaling. The protected source and destinations still cannot be overwritten.
• Instrumental and Acapella exports continue to explicitly refuse any missing, outdated or corrupt stem; no substitution with the original audio is permitted.

• Stems from the same source are now recombined before applying the rack inherited from the clip. Compressors, saturations, limiters, restores and other non-linear processing are no longer performed separately on each stem.
• The recombined group then retains the track, automations, busses and mix from the original clip. Original audio is never added in parallel when valid stems are active.
• The stems window offers a non-destructive Original/Processed comparison. The bypass and all stem settings remain saved in the project and follow Undo/Redo.
• The measurements after summation separately expose the sample peak, the EBU R128 loudness and the true peak. The same graph remains used for listening, analysis and export.
• Phase tests cover non-linear effects after sum, lack of duplicate original/stems, A/B bypass, Solo/Volume, linked audio tracks and measurements after mix.

• Separately outlines the actual decoding, transfer, conversion, composition, presentation and encoding engines.
• Measures the time, volume and throughput of uploads of reused OpenGL textures.
• Explicitly distinguishes the Aesthetic HDR Effect from the HDR10 PQ/BT.2020 output.
• Adds these capabilities and fallback paths to the redacted diagnostic.

• Excludes reconstructable waveform cache from history snapshots.
• Bounds the history both by number of actions and by serialized memory budget.
• Virtualizes already filtered timeline elements and measures their redrawing at p95.
• Maintains a continuous gesture in a single Undo/Redo action.

• Measure RAM, VRAM, GPU, threads, handles, disk and duration of AI operations in a bounded log.
• Serializes AI splits and pauses background worker during interactive playback.
• Reserves the GPU for the active preview by choosing the CPU for a new concurrent split.
• Limits CUDA to CPU fallback to hardware, driver, memory, or CUDA runtime errors.

• Pools manual and automatic writes with single temporary, lock by destination, fsync, atomic replace and previous copy.
• Rejects invalid JSON numbers and only publishes saved state after complete success.
• Limits the wait for the decoder to close without freeing a container still used by its thread.
• The task manager returns the report of tasks or processes that have not finished within the deadline.

• Persists download, verify, extract, qualify, ready, damaged and delete states.
• Keeps the old installation until complete qualification and restores it if the switchover fails.
• Automatically recovers an interrupted installation and cleans up staging folders.
• Coordinates use, installation and removal; Repair and deletion run outside of the UI thread.

• Continuously drains worker output, limits diagnostics and applies inactivity and duration time limits.
• Calculates a complete cancelable hash and versions the cache with the Demucs model, runtime and options.
• Checks the readability, duration, channels and frequency of the four WAVs before reuse.
• Reports missing, outdated or corrupt stems in preview and refuses non-compliant Instrumental/Acapella exports.

• Validates the endhost, HTTP status, ZIP type, size and signature before installing the module.
• Immediately closes the Windows temporary handle and cleans up canceled downloads.
• Adds a compatibility protocol, a real worker diagnosis and a “ready” state only after qualification.
• The installed worker now resolves its model offline itself in its own package.

• Adds a qualification report attributable to each compilation: footprint of sources, versions, environment, packages and ignored tests.
• Explicitly separates simulated, real CPU, real CUDA, and packaged EXE proofs.
• Embeds the current report in the autonomous distribution and keeps one copy per phase in qualification/.

• Demucs, PyTorch CUDA and the HTDemucs model are removed from the main Setup and Portable. Equalizer, central blanking, advanced Karaoke and all other tools remain directly available.
• The Voice Removal/Karaoke window now offers to install the AI ​​module online, install a package already downloaded or delete the module in order to recover disk space.
• The standalone AI module maintains validated four-track separation for Voice, Drums, Bass and Others, CUDA priority, CPU fallback, cache, volumes, Solo/Mute and Instrumental/Acapella exports.
• The download uses HTTPS and the archive is only installed after verifying its SHA-256 fingerprint. Extraction blocks paths exiting the dedicated folder and the installation remains separate from the application.
• Separation settings saved in projects are retained when the module is deleted. A reinstallation allows them to be reused without modifying the source media.
• A separate package StreamerEditorPro_AI_CUDA_v1.zip is produced for release alongside the lightweight application. Local installation allows you to use it before going online.
• New buttons, states and errors are available in all five languages.

• “AI Vocal/Instrumental Split” mode actually runs Demucs offline and produces four tracks: Vocals, Drums, Bass and Others. The media is not sent to any Internet service.
• Compute prioritizes NVIDIA CUDA when PyTorch detects a compatible GPU, then automatically restarts on CPU in case of unavailability or CUDA failure. The interface remains available and the operation can be canceled.
• The HTDemucs model comes in Setup and Portable. Its origin, revision, and SHA-256 fingerprint are checked during construction and at standalone startup.
• Results are cached based on source content and model revision. A separation already calculated is reused without reprocessing; The cache of a saved project is placed in its cache folder.
• Each separate track has Enable, Solo and Volume. The Instrumental and Acapella presets, as well as their dedicated WAV exports, use the same non-destructive mixing.
• Separate tracks automatically reflect cuts, positions, speeds, fades, video relationships, mix tracks and effects from the source. The preview, the Timeline and the final export use the same audio graph.
• Cache settings and paths are saved in the project, restored upon reopening and included in Undo/Redo and autosave. A cache miss returns safely to the original source.
• All phase 3 controls and reports are translated into French, English, Italian, Spanish and German. Tests cover true offline model, CUDA/CPU, cache, persistence, Solo/Mute/Volume, sync, and Instrumental/Acapella exports.

• The Voice Removal/Karaoke Analyze button is now active. The bounded, cancelable analysis measures left/right correlation, center signal share, and vocal band energy over a selection of up to thirty seconds.
• Mono sources, strongly centered, partially centered and stereo/off-center vocals are distinguished. The recommended mode, intensity and frequency range are offered in the interface without modifying the source file or automatically validating the treatment.
• A centered voice offers advanced Karaoke; a partially centered voice suggests central deletion; a mono source uses EQ attenuation; an off-center voice recommends phase 3 AI separation.
• Correlation, center proportion and vocal band proportion are displayed after analysis. The result is saved by source in the project and participates in automatic saving.
• Added residual voice reduction with independent activation and intensity from 0 to 100%. This pass restrainedly attenuates the stereo vocal presence, overdubs, and reverbs left after center cancellation.
• Residual reduction uses exactly the same deterministic graph for listening, A/B comparison and export. It remains non-destructive and follows Ctrl+Z/Ctrl+Y with the other voice suppression settings.
• The project format changes to version 4. Phase 1 projects are automatically migrated with residual reduction disabled, which keeps their sound identical.
• All new checks, reports and results are translated into French, English, Italian, Spanish and German. Tests cover centered voice, side content, mono source, silence, undo, recommendations, residual filter, migration and save/reload.

• In Auto Edit, an “HDR Effect” box and a 0-100% setting are added immediately to the right of Auto Enhance. The default value is 20% and the effect remains disabled until its box is checked.
• The tweak exactly reuses the non-destructive HDR effect from the Video tab. It is applied separately to all created video sequences, retained during "Create another order", saved in the project and ignored for thumbnails and images.
• In the Video tab of the Visual Timeline, Contrast, Saturation, Brightness, Sharpness, HDR Effect and Colors now each have a box. “Apply to selection” only modifies the checked settings and retains all other values ​​specific to each sequence.
• Box labels, video buttons, help and HDR messages are available in French, English, Italian, Spanish and German.
• In the Audio tab of the Visual Timeline, two lists allow you to choose any priority track and any track to attenuate; A1 and A2 are only examples and all the avenues of the project are proposed.
• The “Audio replacement” slider goes from 0 to 100%. It reduces the target track only during priority track blocks; at 100%, the target becomes silent on these ranges and returns to its intact level outside.
• 20ms transitions at block edges prevent clicks. Adjacent blocks are joined together to avoid a rise in volume between two continuous sounds.
• Replacement also works with the music track, remains non-destructive, is saved in the project, follows Undo/Redo and uses the same graph for Timeline playback and final export.
• The “Activate / Mute” button now acts in a single operation on all selected audio files. A mixed selection is fully activated on the first click, then fully muted on the next click; Music track and sounds related to videos are supported with single Undo/Redo step.
• The speaker icon of each audio track reflects the state of all its files: no line when they are all active, yellow line when only some are muted and red line when the entire track is muted.
• Added tests cover HDR default, persistence, video checkboxes, keeping unchecked settings, all track pairs, multiple blocks, 50/100% levels, saving, and actual audio processing.

• In Edit Scene, full screen video mode F11 now stays on the current sequence: when playback reaches its end, it resumes at the beginning of this same sequence.
• Same behavior with native video player, 720p preview and reverse playback. The Previous and Next buttons remain available to voluntarily change sequences.
• The loop reuses the limits and playback engine of Edit Scene in widescreen; no automatic transition to the next sequence is planned in F11.
• F11 tests cover looping on an intermediate scene, last scene, reverse playback, native player and resuming after an end of media.
• The Analysis Normalization option now applies exactly the peak normalization from the Normalize button in the Audio window: the highest peak is brought to 100%, regardless of whether the initial level is too low or too high.
• Automatic normalization covers optional music and embedded audio from all videos actually retained in the edit. The gains are non-destructive, saved in the project and the Video ↔ Audio relationships remain intact.
• Each file used several times is decoded only once. Stereo peaks join the common waveform cache; their calculation maintains GPU priority and signals CPU fallback.
• The wording, help, progress, engine used, result and isolated standardization failures are available in French, English, Italian, Spanish and German.
• Normalization tests cover music, multiple sequences from the same video, gain up and down, shared cache, disabled option and keeping independent tracks.
• In Skin, resizing text with the mouse uses thirty-two magnetic steps within each visible tile. At 1920 × 1080, the adjustment advances by approximately 3.75 pixels: enlarging then reducing large text thus remains very precise, while logos, photos, smileys and movement on the grid retain their existing behavior.

• The 10-band EQ now clearly distinguishes vocal EQ attenuation from true stereo center suppression. Its visible range changes to −24/+12 dB, each band displays its value and returns to 0 dB by double-click; mouse, wheel and keyboard remain usable.
• Added the “Voice Removal / Karaoke” panel with five modes: Off, EQ Attenuation, Center Vocal Removal, Advanced Karaoke and the inactive interface of the future AI Voice/Instrumental separation.
• Center blanking converts the stereo signal to Mid/Side, gradually attenuates the Mid in the chosen vocal band then rebuilds left/right. Bass below the limit and treble above the limit can be retained with smooth transitions.
• Added Soft Voice, Loud Voice, Recommended Karaoke and Maximum Karaoke presets, with audit-compliant intensity, frequencies and corrective curves. The maximum warns of the risk for centered instruments.
• The chain ends, when requested, with dynamic normalization and a transparent limiter at −1 dBFS. Listening and exporting use exactly the same non-destructive graph; the source files remain intact.
• Original/Processed prepares the new graph in parallel and replaces it on the active audio clock without reopening the sound device. Listen continues to use the selected source only and does not launch any video previews.
• A mono source requested in central cancellation automatically switches to EQ attenuation in order to avoid silence. The same control is applied source by source during a multiselection.
• Settings are saved by source in the project and integrated into automatic saving. An application on multiple sources constitutes a single Ctrl+Z/Ctrl+Y action; Tracks linked to a video maintain their synchronization.
• Bypass temporarily disables global audio processing. Reset only returns vocal suppression to its default values ​​without erasing other effects from the source.
• The “Audio Tools” menu now updates its four choices in the active language each time it is opened. All new modes, presets, states, settings and messages are provided in French, English, Italian, Spanish and German.
• The names and unit of the ten creative knobs are resolved in the active language with each redraw. The language change now explicitly triggers this redraw for all ten canvases; their old French labels can no longer remain displayed.
• The equalizer buttons are placed immediately below the sliders. The height is calculated after the actual rendering of the fonts and Windows scaling, with a minimum bottom margin: the labels remain whole without recreating the large empty space observed.
• Added tests cover central attenuation greater than 20 dB, Side retention, crossovers, mono, −1 dBFS ceiling, persistence, listen/export equivalence, multiselect, Undo/Redo, A/B replacement without audio reopening, five languages ​​and interface tightening.

• Media relocation can no longer fail with “NoneType object has no attribute winfo_reqwidth” when a tooltip closes during its position calculation. The tooltip now maintains a stable local reference and properly abandons its display if it has been closed in the meantime.
• A “Normalize” button is added immediately to the right of the Noise Generator in “7 - Audio”. It analyzes the selected source then adjusts its volume non-destructively so that its highest peak reaches 100%, increasing or reducing it according to its initial level.
• Normalization respects the independent gains of the left and right channels, never modifies the source file, is saved in the project and uses the common history Ctrl+Z/Ctrl+Y.
• If the stereo waveform is not yet available, the button builds it in the background and then automatically applies normalization. A silent source is preserved and reported without creating artificial gain.
• 10-band EQ, Audio Tools, Noise Generator and Normalize use a uniform grid: the Audio Tools menu is exactly the same size as the buttons to its left and right.
• The new texts are available in French, English, Italian, Spanish and German. Tests cover raising, lowering, silencing, canceling, command order, dimensions, and tooltip reentrant closure.

• The left and right arrows now move the selected source back or forward five seconds, during playback and pause, without opening the video preview. The input fields retain their usual arrows.
• The Audio bar contains the Edit Scene commands: Apply (Ctrl+Enter), Previous Audio (Ctrl+P), Next Audio (Ctrl+N), Exit (Ctrl+Q), Reduce (Ctrl+M), Cancel (Ctrl+Shift+C) and Save (Ctrl+S). Space, Enter, Start, and End retain their existing transports.
• Double-clicking on the waveform enlarges the visualization and double-clicking again restores the interface. Detailed settings are hidden but transport controls, ±5 second jumps and main actions remain accessible.
• F11 activates an Audio only mode which uses the full width for the waveform. F11 or Esc returns to the previous state; the same buttons and shortcuts remain active in this mode.
• Memorizing separators now verifies that two panels are present: leaving enlarged or Audio only mode can no longer display “sash index 0 out of range” during interface restoration.
• Window geometry and vertical/horizontal positions of separators are captured before magnification and then reapplied immediately upon exit by double-click, Esc or F11. The horizontal line thus returns to its original height instead of remaining at the bottom of the screen.
• The seven transport buttons fit on a single line when the available width allows it and keep the same height as the edit buttons below. The Apply/Exit/Cancel/Minimize/Full Screen Audio commands also use this height. Keyboard help is placed on its own row so that long text no longer distorts the entire row.
• The thirteen editing commands and the marker name field use a compact grid of seven columns: they occupy two lines at normal width, including with Italian labels, and collapse further only if the window becomes narrower.
• The four DJ potentiometers and the six direct processing are grouped together on a single line, all grouped on the left. Equalizer, Audio Tools and Noise Generator follow on the right; the slightly more compact potentiometers retain their existing values ​​and gestures.
• All new buttons and indications are available in French, English, Italian, Spanish and German. The tests cover bounded movements, protected fields, both presentations and complete restoration of the interface.

• Right-clicking on the visual Timeline now opens the editor corresponding to the block in question: Modify scene for a video, Start and end thumbnails for a thumbnail, Duration adjustment for a still photo and Skin editor for text, a smiley, a logo or a photo.
• Existing Audio right-click remains unchanged and continues to open the targeted source in 7 - Audio. Scene detection now correctly uses the visual block index.
• The clicked block becomes the sole target of the editor so that a previous multiple selection does not cause an involuntary modification.
• The five added contextual labels are available in French, English, Italian, Spanish and German.
• Only one StreamerEditorPro application can run per Windows session. A second attempt restores the existing window to the foreground and then stops; internal GPU, HDR and diagnostic processes are not blocked.
• Added tests covering all seven right-click block types and the main instance Windows lock.

• The “Cancel setting” button, which duplicated Ctrl+Z, becomes “Reset” in French, English, Italian, Spanish and German.
• Reset resets gain, activation, left/right channels and their link, fades and their curve, tempo, pitch, formants, EQ, DJ effects, six direct processes, rack and track settings specific to the selected source in a single operation.
• The cuts, silences, position, synchronization and markers are preserved: they constitute the editing of the source and not its sound settings.
• The reset is saved in shared history; Ctrl+Z fully restores the previous settings and Ctrl+Y can redo it.
• Added interface and model test covering full reset, retain edit, and undo reset.

• Added six non-destructive knobs: Fast Compression, Pitch ±12 semitones, Reverb, Noise Cleanup, Radio Effect and Telephone Effect. The settings persist in the project, use Undo/Redo and feed the same graph for listening and export.
• White, pink or brown noise generator with duration and dBFS level, 48 kHz stereo WAV creation then automatic insertion into the project.
• Audio transport aligned to the Timeline: Space for Play/Pause, In for Stop, Start/End for the ends of the source and listening volume shared with the active output.
• Shortcuts limited to Audio and its tools; input fields, lists, cursors and relevant controls retain their usual keys.
• Reinforced FFmpeg closure under Windows with wait for source release, qualified on twenty consecutive starts/stops.
• All new controls, indications and statuses are available in all five languages.

• Enter and Numpad Enter control Play/Pause in Audio and its tools, with the same protected scope as Space; the input fields retain their normal behavior.
• The FFmpeg and FFprobe playback, polling and waveform processes are launched hidden under Windows to avoid very short windows on first load.
• The waveform displays the gain and clipping of each channel. Left/Right handles directly adjust volumes, with switchable linkage, automatic gain display, and shared Undo/Redo history.
• Fade in and fade out handles immediately update interface values ​​and reuse the common non-destructive audio graph for listening and exporting.
• Added a ten-band graphic equalizer with Neutral, Vocal Suppression, Loudness, Rock, House and Classic presets. Vocal removal overrides the stereo center and therefore depends on the source mix.
• The interface regroups the four old panels in “Audio Tools” and adds Distortion, Echo, Fade Level and Robot Voice knobs. All new labels are available in French, English, Italian, Spanish and German.
• Real FFmpeg testing of independent channels, central undo, DJ filters, shortcuts, waveform/fade gestures, history and hiding child processes.

• Playback in Audio listens only to the selected source, with gain, edits, processing and automation; other video sources, music and tracks are not mixed. Changing the source stops its previous listening. The export continues to use the full project mix.
• Bounded PCM stream reusing the existing decoder and audio clock; first sound without prior full WAV export, no decoding or video preview. Buffer limited to approximately two seconds, consistent play/pause/move and stop.
• Measurement on Porto Taverna: first samples on the active output in 0.97 seconds, compared to around ten seconds for the old preparation of the complete mix. Three presses keep a single flow, without cancellation. Expensive effects may still require some initial calculation.
• Indication of isolated listening and opening/reading statuses in the five languages. Added isolation tests by frequencies, effects/gain/silence compared to the same export graph, start without waiting for a 60 second file, limited buffer and audio output errors.
• Integrates vertical lift, Space correction, priority GPU for calculations compatible with displayed CPU fallback and all previous Audio phases.

• Added a vertical scrollbar to the right of the list of project audio sources.
• Reading while preparing now keeps the current task: repeated presses no longer cancel it. Stop retains explicit cancellation.
• Validation with the Porto Taverna project: three invocations of the Play button, a single preparation, VLC Playing player, mute disabled, volume 80%, Audio tab active and video preview hidden; unchanged project file.
• Retains fixes for Space shortcut, canceled preparations without error window and late activation of VLC sound. Adjusts the interface test using a simulated application to the keyboard protection of the real toplevel Tk.
• FFmpeg audio calculations: parallelism of filters limited to two tasks such as bounded video rendering, to avoid CPU overallocation of heavy restorations. The keyboard test waits for the actual mapping of the field before giving it focus.
• Scrolling tests on 40 sources and repeated presses during preparation; five languages ​​and historical shortcuts preserved.

• Space now intercepts the press and release before the Tk buttons: only one command per press, without automatic repetition or activation of the button with the focus.
• Scope limited to Audio and its tool windows; text fields, other tabs and independent windows retain their usual keys. Widgets added late are focus-protected.
• An audio preparation normally canceled ends without an error window; true decoding errors remain reported.
• Selection monitoring no longer stops a preparation before it is read; a slider near the end resumes at the beginning. Audio edits/histories do not automatically launch a video preview.
• Retains VLC mute/volume fix and priority GPU scans with explicit CPU fallback from previous revision; real keyboard, reader and GPU calculation validations.

• Fix muted playback: VLC reverted to its old mute/volume state after asynchronous opening. Audio monitoring unmutes the sound and applies volume when the output is actually ready.
• Stereo waveforms, spectrogram and FFT of the noise profile calculated primarily on OpenGL 4.3 GPU; Explicit CPU fallback in case of unavailability or failure, without loss of work.
• Shows the actual GPU used or the reason for the fallback in all five languages; FFmpeg decoding and audio effects remain explicitly CPU-based, their engine not having compatible GPU processing.
• Hann 2048 FFT spectrogram, combination of stereo powers, logarithmic scale 20 Hz–24 kHz and range -120 to 0 dBFS; sources and export processing unchanged.
• Real VLC check of mute, volume, pause/seek/resume status; GPU/CPU compliance, rollback, GPU loss, batch memory limits, and metrics published with no promise of universal gain.

• Playback from Audio and its mixer without opening or activating the video preview; shared transport retained, playback of the audio mix alone with all processing.
• Internal timeline created hidden, without deferred maximization; no window presentation when moving and Audio loops.
• Temporary mix prepared in background, cache invalidated by source/processing, cursor resume, pause/stop/loop and discard late or obsolete results.
• Normal video playback retained when requested from the Timeline; non-regression tests and contextualized controls preserved in five languages.

• Versioned aux busses, pre/post-fader sends, gain/mute/solo returns and multi-tap delay/reverb effects; routing validated without cycles.
• Automatic or manual BPM grid, configurable first beat, indicative confidence, sections aligned with times and persistence in the shared project.
• Explicit use of the grid by the montage generator and its existing adaptation workflow; invalidation after temporal modification or change of source.
• Common continuous playback/export, configurable audio-only queues, backup/Undo-Redo and contextualized controls in five languages.
• Eight real targeted tests; qualification of two identical thirty-minute renditions with sixteen stereo tracks, two busses and thirty-two sends.

• Independent tempo (0.25–4), independent pitch (±24 semitones), optional formant preservation and stereo linked with supplied Rubber Band.
• Non-destructive FFT time/frequency attenuations, cumulative, persistent, editable and cancelable zones; correct source geometry with tempo and reverse playback.
• Atomic batches: separate clips over full duration, full mix and manifest; cancellation without partial file and refusal of existing destinations.
• Complete CPU/GPU read preparation for transforms and FFT zones; continuous video export without restarting processing at each block.
• Neutral gain of the optimized mixer; protection of the last samples, compensation and final length imposed at 48 kHz in number of samples.
• Qualification of two identical exports of thirty minutes, sixteen stereo tracks with EQ/compressor; real tests, five languages ​​and shortcut controls.

• Selection spectrogram; mix analysis with LUFS-I, True Peak, Loudness Range, peak, RMS and crest factor.
• Normalization analyzed, configurable targets and choice of limitation/reduced target; measurement of the final result and refusal of outdated analyses.
• Persistent capture of noise profiles, FFT reduction with 25ms compensation, DeHum/harmonics, DeClick, DeClip, DeEsser and high pass.
• Camera, drone, voice, male/female, interview, narration and podcast presets; Editable channels in racks.
• A/B comparison of effects, Undo/Redo and saving of the shared project.
• Cancelable temporary mix preparation for clipping repair before playback; cache invalidated by the source, effects and normalization target, retaining the same processing on CPU/GPU/export.
• Translations in five languages ​​and real test corpus: noise, hum, click, saturation, sibilants, silence, loudness, spectrogram, cache and persistence.

• Mix tracks from the shared project: gain, stereo balance, mute, cumulative solos and grouping of clips by track.
• Automation of volume and balance in project time, interpolation by sample, saving and Undo/Redo.
• Clip/track/master effects chains: Parametric EQ, compressor, gate and limiter with versioned parameters, bypass and reordering.
• Compensated limiter anticipation; Bounded preloading of processing during reading moves.
• Continuous audio mixing for video export with effects, avoiding compressor restarts at video block boundaries.
• Effects and markers preserved during copies, collages and cuts; independent effect identifiers.
• Contextual interface and shortcuts in five languages; new DSP tests, automation, copy, history and video export on several blocks.

• Non-destructive editing: copy, cut and delete while leaving a gap, paste, duplicate, silence, trim, crop, cut and explicitly uncouple videos.
• Time selection at sample precision, slider, loop over selection, zoom, view pan, left/right/stereo channels and renamable markers.
• Audio or video import with sound, audio-only sessions, resizable panels and view state saved in the shared project.
• Gain, activation and four fade curves; a validated gesture constitutes an Undo/Redo command.
• Video export corrected to keep manual gain in final assembly; preview and WAV use the same graph.
• Translations into five languages; real tests of DSP, persistence, cache, undo, keyboard and video/audio sync after seek, speed and reverse.

• CTRL+S (Save) and CTRL+Y (Redo) indications in all five Audio Studio languages, reusing existing shortcuts.
• New main tab 7 - Audio in all five languages, to the right of SRT subtitles.
• Shared project sources, music and linked video audio; Contextual access from timelines.
• Non-destructive gain -60/+24 dB, enable/mute, shared history, existing video save and export.
• Cached source waveform, transport reusing the visual Timeline and its synchronized preview.
• Persistent manual mixing with peak protection; 24-bit/48kHz stereo WAV export using the shared, atomically verified and finalized audio graph.
• Audio Studio delivered in phases: advanced effects, mastering and additional audio formats remain to be developed.

• In the Timeline tab, Move Up now uses Ctrl+U, Move Down Ctrl+D and Duplicate Ctrl+Shift+D.
• The corresponding indications are updated in French, English, Italian, Spanish and German.

• In Edit Scene, Delete removes the current video from normal view, wide screen and F11 video mode; input fields retain their usual delete key.
• In the Timeline tab, Del, Ctrl+A, Ctrl+E, Ctrl+U, Ctrl+D and Ctrl+Shift+D execute Delete, Select All, Edit, Move Up, Move Down and Duplicate respectively.
• The old labels To the left and To the right become Up and Down.
• The six Timeline controls and scene deletion have indications translated into French, English, Italian, Spanish and German.

• Ctrl+1, Ctrl+2, Ctrl+4, Ctrl+8 and Ctrl+0 select preview speeds ×1, ×2, ×4, ×8 and ×10 respectively.
• Ctrl+O returns playback to editing speed.
• Shortcuts remain active in widescreen and F11 video modes without changing the speed recorded in the scene.
• The indications and tooltips of the six commands are translated into French, English, Italian, Spanish and German.

• Reverse video sequences now display a purple R badge in the Visual Timeline.
• The badge is refreshed immediately after the Reverse video, Ctrl+R, Undo or Redo button.
• Still images do not receive this mark.

• After a deletion, the adaptation to the music first uses the drone passages already present before adding a new sequence.
• A small reduction in speed creates the necessary margin, then the cut remains strictly within the analyzed smooth passage: the end cursor moves to the left first, or the start cursor to the right if the end is locked.
• The deficit is distributed between several drone videos in order to limit changes in speed and cutting on each pass.
• The source limits of a drone are never extended during this operation; cursor locks are respected.
• Precise drone resynchronization uses the same order of priority for its cursor corrections.

• Ctrl+R reverses the video in Edit Scene and in the Visual Timeline, including in full screen views; the shortcut appears on the buttons in all five languages.
• Reinstallation after an update restores missing settings and persistent files once, without overwriting the user's more recent choices.
• The Auto Enhance option is enabled by default in Auto Edit and saved in the project.
• AI-free local generation of 5-second start and end thumbnails is enabled by default, saved to the project and maintains music duration by adjusting timeline content.
• Shorts Studio's Intro and Out editors support Ctrl+A, Ctrl+X, Ctrl+C, Ctrl+V, Ctrl+Z and Ctrl+Y on their graphical elements, while retaining native shortcuts for text fields.
• The Skin editor accepts Ctrl+A and clipboard operations apply to all selected elements.

• After modifying the durations or composition of the sequences, closing the visual Timeline suggests adapting the montage to the duration of the active music.
• Yes opens the existing proposal; the Timeline closes after application. No closes the window while retaining the changes.
• Cancel the proposal keeps the Timeline open. A preparation error also leaves the assembly available.
• No message for a simple image inversion or correction, modification canceled, or music disabled.
• Question translated into five languages.

• In widescreen Edit Scene, the two save lines appear on the right in Scene Settings, aligned to the left.
• In F11, the two lines are placed to the left of Next, without additional window base.
• The F11 bar offers Save (CTRL+S), translated into five languages; the same shortcut remains active in both modes.

• In Edit Scene, the two save lines are placed to the right of the Smooth Flow Control frame, with the text aligned to the left.
• Removed the window footer reserved for saves in order to recover the height for the preview and commands.

• The 18 Media and Editing controls are distributed over three rows of six buttons when width allows.
• Removed the imposed line break which created two rows with a single button.
• The bar continues to adapt to narrow windows and Windows scaling.

• The Select Sequences label fits on two lines in all five languages.
• Save button to the left of Reverse Video, translated into five languages.
• Ctrl+S saves the current project and scene settings, including F11.
• Display at the bottom right of the project recording and automatic save times.
• Existing autosave incorporates valid scene editor settings; an invalid entry defers saving without interrupting playback with a message.

• SRT progress remains visible at the top of the tab, with percentage and current step; the forms scroll separately.
• SRT editor with listening, waveform, beat adjustment and style preview on a video frame; Saveable styles and embeddable subtitles for export.
• Lyrics synchronization improved by reconciling the recognized text and the lyrics provided; the estimated times remain verifiable in the editor.
• Control of media, image-free intervals and subtitles before export.
• Conservation of completed export blocks to resume an interrupted rendering with the same sources and settings.
• Smooth preview cache management and saving/restoring edit variants.

• Two menus allow you to choose each language of the duo separately from French, English, Italian, Spanish and German.
• The “SRT Subtitles” tab accepts a single bilingual TXT file with the markers [FR], [EN], [IT], [ES], [DE] and their combinations.
• Whisper only identifies sung moments; the displayed texts come exactly from the TXT file and follow the actual order of the languages ​​in the song.
• During editing analysis, the only TXT of the music folder is automatically associated with the MP3 and creates its bilingual SRT subtitles.
• Section titles such as “Verse”, “Refrain”, “Strofa” and “Ritornello” are ignored during synchronization.

• A new “SRT Subtitles” tab transforms an MP3 file into SRT with Whisper, entirely locally.
• Language can be detected automatically or chosen from French, English, Italian, Spanish and German.
• Actual Whisper progress is displayed and transcription can be canceled without producing an incomplete file.
• During “Analyze and Create Montage”, the main music automatically receives an SRT file with the same name.
• A musical SRT already present is kept without modification; Instrumental music or impossible transcription does not block the creation of the montage.

• The “Analyze and Create Montage” window displays the elapsed time and an estimate of the time remaining for the entire processing.
• The estimate is recalculated from the actual progress and continues to update between two files.
• If a video with the same name exists and its replacement is refused, the export starts from the project name and uses the first available suffix (Project name_01.mp4, _02, etc.).
• Duration indications are available in French, English, Italian, Spanish and German.

• Create another order calculates the editing in the background: the window remains usable during preparation.
• The selection of passages avoids repeated checks on each file, which greatly reduces waiting on long projects.
• The selected passages now fully cover each planned cut; the duration of the timeline remains the same in the three camera/drone modes.
• Verification on a real project of 62 minutes: new order created in a few seconds, with a final duration of 3,721.844 seconds.

• GPU preview audio preparation divides long projects into portions of limited duration: Windows no longer receives an FFmpeg command containing the paths of all the scenes in the montage.
• The size of each order is checked before it is launched and the portion is reduced if necessary. Audio samples remain aligned from one portion to the next.
• Verification on a real project of 62 minutes with 682 sequences and 156 audio tracks: the preview starts with images decoded by CUDA and sound, without Windows error 206.

• The Reverse Video button now also reverses the preview playback in Edit Scene, including in F11. Playback uses the frames from the smooth preview; its preparation is expected if necessary, without launching the video in the normal direction in the meantime.
• The cursor starts from the end of the reversed passage, loops to the beginning in the normal window and ends the last sequence in F11.
• Exit, close with the cross or Esc, as well as Cancel, ask what to do with unapplied changes, including reversing. The request allows you to apply, abort, or stay in the window.
• The closing question and the Apply and close button are available in all five languages.

• The final export progress window displays the video name, duration, estimated file size, format, codec, resolution, frame rate and bitrate.
• The color space remains deliberately absent from this panel.
• The estimated size takes into account the chosen video bitrate, AAC audio and a small margin for the MP4 container.
• The information is frozen at launch in order to correspond to the project actually transmitted to the rendering engine.
• The new labels and units are translated into French, English, Italian, Spanish and German.

• Create another order now uses the actual timeline duration before the action, even if the target duration field still contains a longer value.
• The new order is only applied if the final duration remains identical to the thousandth of a second; otherwise the existing assembly is retained in its entirety.
• Regeneration is atomic: a failure no longer modifies the scenes, the random seed or the project settings.
• Camera then drone, drone then camera and random camera/drone mix modes are all controlled with this fixed duration.

• Added Invert Video button to the left of Full Screen Video (F11) in the Edit Scene window.
• The wording includes the French, English, Italian, Spanish and German translations of the visual timeline.
• The inversion acts on the current sequence and is recorded by Apply, with the same editing property as in the Video tab of the visual timeline.
• An unapplied inversion is recognized as a pending modification when moving to the previous or next sequence.

• Ctrl+M now works as a toggle in Analysis, Export, Edit Scene and Skins.
• After minimizing, the same shortcut restores the window to the foreground even if it no longer has focus.
• The temporary shortcut is released after restoring or closing the window.
• Skins also restores its modal lock after reopening.
• The message reporting a conflict with another application is translated into all five languages.

• The build suite always uses canonical French for its message comparisons.
• User chosen language in StreamerEditorPro can no longer cause false test failures.
• Personal language preference remains unchanged and explicit testing of all five languages ​​continues to work.
• Validation: the 13 previously failed modules pass, then the 164 construction modules are validated.

• The Export tab offers a list of 13 speeds, from 5,000 to 100,000 Kbits/s.
• A new project uses 25,000 Kbps by default.
• The chosen bitrate is applied to encoding, saved in the project and restored when it is reopened.
• Publication profiles continue to apply their recommended bitrate and the choice is also used by batch export.
• The setting wording is translated into French, English, Italian, Spanish and German.

• Ctrl+M now minimizes the export window then restores it with the same shortcut.
• In Windows, the shortcut remains available while the window is minimized and no longer has focus.
• The temporary shortcut is automatically released after restoring or closing the export.
• Validation: reduction/restore cycle verified on a real Windows window, with dedicated unit tests.

• The progress window is restored, elevated and receives focus as soon as an export or preview creation is launched.
• A temporary priority of 350 ms guarantees its passage in front of other windows under Windows.
• The priority is removed automatically: the export window does not always stay on top during rendering.
• Validation: real focus, normal state and automatic removal of priority verified under Windows, with dedicated unit tests.

• The version is now presented only as 0.20.175.
• The internal update channel remains separate from the version number and continues to direct beta updates without appearing as a suffix.
• Updated version notices in the five installer licenses.

• New multi-resolution Windows icon generated from the 1254×1254 source logo, with real transparency.
• Crop transparent margins so that the logo uses almost all the available area in the taskbar.
• Sharpness and contrast adapted separately to sizes 16, 24, 32, 48 and 64 pixels; preservation of HD details in 128 and 256 pixels.
• Added a dedicated 512x512 transparent PNG to the preview windows and a reproducible brand assets generator.
• Validation: seven resolutions checked, useful padding checked, native Windows icons active and 23 targeted tests passed.

• Explicit restoration of the native Windows .ico icon in the title bar of the main window and secondary windows.
• Preserve multi-resolution images for high-density screens and taskbar.
• Reapply the icon after initial display so that Windows can no longer replace it with the default Tk pen.
• Validation: control of the seven icon resolutions, native Windows calls and the actual icon handle after display.

• Fixed the “SimpleNamespace object has no attribute _reset_timeline_edit_state” error encountered by the Windows Builder project opening tests.
• Resetting Timeline history now also works with simulated objects used by tests, without changing the behavior of the real application.
• Updated logo boot checks and 720p GPU previews so that they check the paths actually used by the current version.
• Validation: 77 targeted tests passed, including opening, recovery, new project, Timeline history, startup and smooth previews.

• Enables Ctrl+Z, Ctrl+Y, Ctrl+X, Ctrl+C, and Ctrl+V throughout the Skins window, including when the focus is on a button, list, or preview area.
• Copy retains all properties of the selected skin. Paste creates an independent element, selects it, and shifts it one grid square to make it immediately visible.
• Cut and paste contribute to undo and redo history. The history remains limited to 30 steps.
• Help and status messages available in French, English, Italian, Spanish and German.
• Validation: tests dedicated to the clipboard, the five keyboard connections, the history and the translations.

• Activation of Ctrl+Z, Ctrl+Y, Ctrl+X, Ctrl+C and Ctrl+V in the main Timeline tab, including when the focus is on a button in this tab.
• Copy and cut take all selected sequences. Paste inserts after the last selected sequence, creates new identities and selects the result.
• Linked audio tracks and transitions within the selection follow copy, cut, paste, undo and redo.
• Ctrl+Z and Ctrl+Y also cover moves, duplications, modifications and deletions made in this list. History limited to 30 steps and reset when changing projects or opening the visual timeline.
• Shortcuts do not capture input in text fields and remain limited to the Timeline tab. Messages translated into five languages.
• Validation: 124 targeted tests passed and real test of the five shortcuts in the main window passed.

• When playing F11 of Edit scene, the intermediate sequences continue to follow each other automatically, then playback stops at the end of the last sequence.
• Removed loop replay of the last passage, both with smooth preview and with the native VLC player. Any return request already prepared by the reader is canceled.
• The last image remains displayed. A new Play command restarts the last sequence from its beginning.
• Outside of F11, loop playback of the selected passage is retained.
• Ending message available in five languages. Validation: 144 targeted tests passed.

• Removal of the volume banner added in 0.20.166 in the Tk preview window.
• Conservation of the volume of the OpenGL window and its synchronization with the listening volume of the timeline. Strict automatic GPU mode is retained.
• Validation: 39 targeted tests passed.

• Added the 0–100% volume slider in the banner of the GPU Preview window (OpenGL), separate from the Tk window. Click and drag; synchronization with the listening volume of the timeline, without modifying the project or the export.
• Auto GPU becomes the default choice for this preview. Hardware decoder testing available following Intel/NVIDIA preference; no last CPU try in this mode. Windows GPU numbers are not used as fixed identifiers.
• After an incident, resuming the process maintains GPU mode instead of imposing CPU. Temporary exclusions of an old test do not deprive strict mode of new hardware tests. Limited number of times.
• OpenGL paths without compatible GPU conversion (including some HDR/transitions) stop with a message. CPU mode remains available by explicit choice. Audio processing and certain conversions remain on CPU.
• Validation: 115 successful targeted tests, including simulated NVIDIA to Intel fallback and CPU refusal after failure of all GPUs. Real test of OpenGL window on H.264 video: CUDA active, 33 frames presented, none lost at measurement point, no errors. User media remains to be verified.

• Removed slider and volume percentage in F11 mode of Edit Scene; timeline and playback controls retained.
• Added listening volume from 0 to 100%, with percentage, under the video in the Preview window. Value shared with the timeline listening volume.
• The setting affects the drives only, without modifying the project or export volumes. Wording available in all five languages.
• Validation: 81 targeted tests passed (volume, F11 commands, reading and translations).

• New Video Acceleration setting: Automatic, NVIDIA priority or Intel priority for creating smooth previews, rendered previews and exporting the edit.
• Capacities verified by real tests; manual priority with fallback to other compatible paths then the CPU. The encoders explicitly imposed on export remain respected.
• Preference associated with the name of the computer, distinct from the project; return to Automatic if settings are transferred to another PC. Abilities are not remembered between casts.
• Preserve existing previews and other preferences. Interface translated into five languages, confirmation after Apply.
• VLC playback: the choice of card remains controlled by Windows; access to Windows graphics settings with restart indication. No forced card switching during playback.
• Validation: 123 targeted tests passed. Real tests on this PC: Intel QSV preview with CPU fallback decoding, NVIDIA CUDA/NVENC preview, Intel and NVIDIA H.264 export profiles verified at 1280x720. No gain in fluidity or speed guaranteed by these tests.

• Direct reading of existing preview files by VLC, in Edit Scene in normal mode like in F11, instead of PyAV positioning and conversion to Tk image for each image.
• Mouse navigation preserves movement images; the pause returns to the original source at the same time. No change to cache format and no regeneration required.
• A ready cache becomes priority for reading. If VLC is unavailable, the image player remains a fallback; without cache, the source remains usable.
• Validation: 156 targeted tests passed. Real VLC test on a pre-existing cache: 72 images displayed / 0 lost in normal mode, 63 / 0 in full screen over two 2.5 second measurement windows including startup.
• These measurements concern the test video; performance on user media remains dependent on the hardware and speeds chosen.

• Ctrl+M restores the minimized window, returning to its previous layout, including F11.
• On Windows, Ctrl+M reservation only during reduction. Release upon restoration or closure; auto repeat disabled.
• Report a conflict if another application already reserves this shortcut.
• Validation: 30 targeted tests passed.

• Removed the generation of a continuous file from the montage to the input in F11 and the associated controller.
• Priority to smooth 720p previews already ready in F11; original reading while waiting for an absent insight. Opening F11 no longer triggers a full render.
• Automatic transition to the next scene at the end of the passage, with retention of confirmations and Next/Previous commands. The last scene keeps its playback loop.
• Preservation of the passage timeline, F11 buttons and volume of the original player. Existing smooth previews remain without sound.
• Validation: 152 successful targeted tests and real testing of two scenes using their pre-existing caches, without requiring proxy generation or editing rendering.

• The choice to delete when closing is stored in the application preferences, independently of saving the project. An old project no longer replaces this choice.
• When retention is chosen, automatic cleaning according to the 4 GB quota no longer deletes inactive previews from the application cache. The disk space check before creation remains active.
• Invalid caches or modified sources remain rebuilt. Explicitly removing a source retains its existing behavior.
• Validation: 10 targeted tests passed, including real reuse of an overview and intact source, retention beyond quota and restoration from an old project. The image comparison test uses the deterministic CPU backend.

• Preparation in the background of a single 720p preview of the scenes in the window, in their order, with clean cuts. The project and its export transitions are not modified.
• In natural reading, transition between scenes without changing player or reopening sources. Next/Previous moves playback in the same file while maintaining its play/pause state.
• Timeline of the selected passage, synchronized cursor and source time, Previous, Play/Pause, Stop and Next buttons accessible in F11, independent listening volume.
• Confirmation of changes retained. Passage changes invalidate the rendering and trigger its preparation. Return F11 to the original source with preservation of unapplied values.
• Progress and errors translated into five languages; canceling jobs and cleaning temporary files on shutdown.
• Validation: 154 targeted tests passed. Real tests with FFmpeg and VLC on two files: red/blue cut verified in the encoded images, natural sequence, same player preserved by Next/Previous, Stop, timeline and conservation of modifications at F11 output.
• Entry into F11 requires preparation of the preview; a manual movement remains subject to the VLC positioning time.

• Additional reduction of 20%: width and height to 64% of the original size, with proportions, centering and transparency preserved.
• Validation: 4 successful startup tests.

• Reduced the width and height of the logo displayed at startup by 20%, maintaining proportions, centering and transparency.
• Validation: 4 successful startup tests.

• The PNG already has a transparent alpha channel outside the logo. The window and its image component, however, imposed a dark rectangular background.
• Activation of the transparent color of the Windows window to let the desktop appear around the logo.
• Validation: 10 targeted tests passed, including verification of native transparent color under Windows.

• Replaced the startup logo with a cut-out version in transparent PNG with the integrated editing tool.
• Background of the display component harmonized with the dark background of the home screen.
• Validation: alpha channel present and 10 targeted tests passed.

• Centered display of the logo provided, proportions preserved, up to 760 pixels and limited to 70% of the screen.
• Logo visible before loading the interface modules, for at least 1.5 seconds unless closed by click or Esc.
• Original image included in application resources and Windows build; diagnostic modes and auxiliary readers remain without a home screen.
• A missing or unreadable image does not block the launch. Projects opened from Explorer are always transmitted to the application.
• Validation: 40 targeted tests passed.

• When the next or previous scene uses the same source and its player is operational, keep the VLC player, decoder, metadata and interface instead of destroying them and then recreating them.
• Direct positioning at the new start; resetting old move requests and temporary markers, updating duration, speed, locks and navigation buttons.
• F11 mode and listening volume are retained. Resumption of reading follows the existing behavior.
• Another source, a reader in error or a comparative measurement in progress retains the complete loading path.
• Validation: 142 targeted tests passed. Unencrypted time saving on user media.

• Logs show images displayed in progress while the position remains at the start of the passage, followed by a recovery after five seconds.
• Positioning confirmation in continuous reading now accepts the time traveled since the positioning request, with a new displayed image required. The cursor can thus follow real time without waiting for a restart.
• Precise position control remains unchanged for paused moves and skip frame fallback mode.
• Regressions covered: first late display at speeds 1, 3 and 8, progress after confirmation, rejection of an inconsistent time and paused positioning.
• Validation: 140 targeted tests passed.

• In F11 mode, Next and Previous restart playback at the start of the new passage if it was active before the scene change.
• The reading intention is preserved before the modification confirmation pause; resumption occurs after validated navigation only.
• A paused preview remains paused. An out-of-bounds or canceled navigation does not launch the new player.
• Validation: 138 targeted tests passed.

• Defect reproduced with real Tk fields: their creation rewrites the displayed duration to six decimal places and triggers a conversion to the source duration, creating a modification without entry.
• A rewrite identical to the displayed duration now retains the precision of the source duration. Genuine entries remain taken into account and subject to confirmation.
• Regression test with the two duration fields, at speeds 1, 3 and 100: precision preserved, navigation without alert, real modification detected.
• Validation: 135 targeted tests passed.

• Correction of the perimeter of the timeline F11: it only represents the passage between the start and end marks, and follows their modifications.
• Cursor movement limited to this passage; display of the time elapsed since its start and its duration in source time, with green and orange marks.
• Reading limited to the passage also selected in F11. No change of cutting points by the reading bar.
• Listening volume independent of editing and export retained, as well as keyboard shortcuts.
• Validation: 134 targeted tests passed.

• Compact bar across entire source video with draggable playback slider and current time/total duration in hours, minutes, seconds and milliseconds.
• Playback in F11 on the entire source, without changing the cut points of the scene. Return to scene playback behavior when exiting F11.
• Listening volume from 0 to 100%, reserved for the player, kept between scenes and when the player restarts, without modifying the edit or the exported volume.
• Wording translated into all five languages. Keyboard shortcuts remain available.
• Validation: 133 targeted tests passed (navigation, volume, reading, shortcuts, translations and window restoration).

• Scene controls remain accessible from the keyboard when their buttons are hidden by F11.
• Disabled buttons and commands outside this overview remain excluded. The Ctrl+K palette includes hidden scene controls.
• Previous/Next navigation and scene reloading retains video mode only, with confirmation of unapplied changes.
• Tk keyboard tests with Num Lock: apply, navigate, cancel, exit, minimize, recalculate, play/pause and Stop.

• Fault reproduced: the keyboard mask 0x8 of Tk under Windows represents Num Lock, but the filter treated it as Alt and ignored Enter/Space.
• Fixed the filter depending on the platform in Edit scene and timeline. Alt and Ctrl remain distinct from keyboard locks.
• Tk tests with Num Lock, Caps Lock and Scroll Lock; Ctrl/Alt controls retained.
• Tk 8.6 reference: https://github.com/tcltk/tk/blob/core-8-6-branch/win/tkWinPointer.c

• Priority input and digital input in the timeline and its preview, including buttons, fields and sliders.
• Modify the scene: installation of the shortcut on the controls created after opening, when they gain focus.
• Ctrl+Enter and separate dialog windows remain independent of Stop.
• Real Tk tests of keyboard propagation and single trigger.

• Full screen video (F11) placed to the left of Next on the same line.
• Removed empty lines between navigation and timeline.
• Apply confirmation integrated into the central text, without additional height.

• Full screen video button (F11): hides settings, navigation and controls, without recreating the player. Escape/F11 or back button restores previous mode.
• Before Previous/Next: if the scene has changed, suggest Apply and continue, Abandon changes or Stay on this scene.
• Comparison with the last applied state, including smooth transitions, locks and invalid fields.
• Texts available in FR/EN/IT/ES/DE.

• Use the previous passage button placed to the left of the next one, in normal and full screen display.
• FR/EN/IT/ES/DE translations, cyclical reverse traversal, restoration of cached passages and respect of locks.
• Scene duration preserved; controls spread over two lines to avoid hiding the buttons.

• Green banner without dialog box after successful application in scenes, skins, transitions, video settings, adaptation and Shorts templates.
• Validation confirmation in the parent window of closed dialogs.
• The banner remains for 4.5 seconds, can be hidden by clicking and remains independent of the reading messages.
• Already applied video settings are reported without creating an additional operation.

• Persistent, time-stamped green confirmation in Edit Scene after successful application.
• A new setting indicates that changes remain to be applied.
• The reader can no longer replace this confirmation with its reading state.

• Space: play/pause and Enter: stop/return to start, with priority over window controls.
• Shortcuts preserved after navigation between scenes and in full screen. Ctrl+Enter keeps Apply.
• Tk keyboard propagation tests, absence of double triggering and window isolation.

• Choice of camera then drone, drone then camera, or random mixture before analysis; backup and regeneration.
• Limited recovery of the active GPU decoder after an incident; temporary exclusions instead of permanent ones.
• Validation of the order and musical adaptation compatible with the blocks.

• Video tab: Invert selection becomes Invert video.
• Media and editing: new line under existing actions with Select All (Ctrl+A), Deselect All (Esc), Cut (Ctrl+X), Copy (Ctrl+C), Paste (Ctrl+V). Common dimensions, line wrap on narrow window.
• Select all keeps start/end videos and thumbnails; Deselecting all clears the selection of all types.
• Internal clipboard for mixed video, image, skin and audio selections. Independent copies with new IDs, settings/fades retained. Cutting leaves the source files intact and preserves the positions of the remaining clips.
• Paste to cursor, postponed to the end of the traversed video clip if necessary; following clips shifted to the affected tracks. Thumbnail copies become regular images. Ctrl+Z cancels cut/paste.
• Shortcuts for text fields preserved; Esc keeps full screen output.
• Validation: 31 targeted tests passed. Windows display and shortcuts to be confirmed.

• The .sepro.json file also contains the state of Shorts Studio: sources and cleaning, maps/images and elements, durations, music, output, grid/snap settings and generation choices.
• Save to Shorts Built-in Studio calls the main project save. Old Shorts JSON files can still be imported by their load command.
• Restore when opening the tab; Closing and replacing the project detects changes specific to Shorts Studio. The workers in the old tab are stopped when changing projects; change blocked during active export.
• API keys and runtime objects are not saved. Media remains referenced by their path, as in the main timeline.
• Targeted validation: 38 tests passed, 6 conditional tests ignored; round trip of the common project, Shorts modification only, saving from both interfaces and compatibility of old projects.

• Export of the Shorts Studio montage: H.264 NVIDIA NVENC test in real vertical format, then GPU encoding of the shots and in/out cards; final assembly without video re-encoding.
• Source: NVDEC test and explicit download to CPU before cleaning, cropping and effects. Automatic CPU fallback in case of unavailability, with reason displayed and stored in manifest.json.
• Preview cleaning and navigation: priority GPU decoding; creation of preview copies via NVENC when available. Card composition, effects and Tk/Pillow display remain CPU.
• Separate display of encoder and decoder. Bounded and cancelable hardware tests; Preview work interrupted at batch launch.
• Validation: 30 tests passed, 1 graphics test ignored, including real music/cards/cleaning exports and time limit; Navigation overview tested from start to finish. RTX 4060 to be confirmed on the Acer, no numerical gain announced.

• Auto Enhance now applies an HDR effect of 20% instead of 30% to selected sequences.

• Texts, logos, smileys and wrapping images: turquoise/orange fade handles at the top of the clip, duration adjustable with the mouse and Ctrl+Z cancellation.
• Transparency envelope saved, limited to the duration and kept after navigation or cutting of the rendering. Adapted OpenGL previews, CPU/FFmpeg and native composition. Video frames retain their existing fades.
• New Colors slider (%): 0 neutral, 100 automatic recommendation, applied only to selected sequences. Dose contrast, saturation and brightness, without modifying HDR/sharpness. Auto Improvement sets it to 100; reset to 0.
• Transparency control on OpenGL pixels and FFmpeg segmented text/logo exports; selection and persistence tests. Continued before adding Colors: 258 passed, 9 ignored; last targeted check: 22 passed.
• Windows graphics handling to be confirmed. The pre-existing limitations of the dedicated HDR10 path regarding skins remain applicable.

• Compares the project and setting fields with the status of the last successful save, open or creation.
• Closure without new request if no modification; confirmation retained for unsaved changes, even if still entered in a field.
• A canceled or failed save does not mark the project as saved; rolling back changes to the saved state removes unnecessary request.
• Eight targeted tests passed.

• Non-modal rendering progress, without forced foreground: the main window and the timeline remain accessible and can be minimized.
• Reduce button in progress; encoding and counting continue in the background. Cancel rendering retains its action.

• Progress window: elapsed time and estimated remaining time, updated twice per second.
• Estimate of the observed progression; recalculation when moving to final assembly and indication of calculation in progress when data is insufficient or progress is blocked.
• Stopping the countdown upon cancellation; zero at the end. Reserved undo bar at the bottom, enlarged window.
• Nine targeted tests passed; Windows display to be confirmed.

• New project restores the original title.

• Close and Insert/Replace use the same width in the fixed bar.

• Reserved Insert/Replace and Close bar at the bottom before the expandable gallery: proposals can no longer hide actions.
• Status messages on a separate line from the buttons; AI panel placed before the gallery when activated.

• The real profile NVENC test no longer depends on a first 128 × 128 miniature test that could prematurely reject the hardware.
• NVENC detection in the FFmpeg actually used, test at requested format/cadence/depth and GPU priority, including for old libx264/libx265 projects.
• CPU fallback of the same codec only after absence of the encoder or real failure; FFmpeg cause displayed and logged. NVIDIA imposed remains an error on failure.
• Preliminary test used by other paths increased to 1920 × 1080 and 20 second delay.
• Automated validation of hardware selection and software exports; confirmation on the RTX 4060 still necessary.

• Select All includes video clips and start/end thumbnails present; other photos, audio tracks and graphics remain excluded.

• Two buttons to the right of Invert Selection: Select All (video clips only, no photos/audio/skins) and Deselect All (all item types).
• Buttons of the same size, labeled on two lines to remain readable in a reduced window.
• The selection does not modify the editing or playback; Pending automatic settings are canceled when changing selections.

• Sequential export by blocks; temporal selection before inversion, memory budget of inversions and limitation of simultaneous sources.
• Supervised cancellation independently of FFmpeg, bounded stderr, memory monitoring and delay outputs; suspending timeline previews.
• Verification of the real NVENC profile and NVDEC sources; encoder and CPU fallback displayed, old CPU choices explained.
• Video assembly without additional re-encoding; global normalization and fix for zero duration fades.
• Existing file replaced only after success; local resource diagnostics, optional GPU via nvidia-smi.
• experimental scale_cuda disabled by default, awaiting comparison on the Acer. Guide: VALIDATION_EXPORT_UHD.md.
• Validation: 244 tests passed, 9 ignored depending on the capabilities of the environment; real SDR/HDR exports, inversions, holes and overlays.
• Player Edit Scene unchanged. No quantified speedup or hardware reboot resolution claimed.

• Commands retained at startup, states and first presentation confirmed, monitoring of blockages and limited recovery.
• Isolated and supervised OpenGL and HDR10 players; bounded closure of preview processes.
• Grouped moves, preparation of the next source, bounded inverted cache and jump strategy according to keyframes.
• Avoided redundant OpenGL transfers, volume conversion outside audio lock; Slower buffer variant discarded after measurement.
• Local diagnostics without media and Windows SDR/HDR10 comparator: compare_apercu_windows.bat. Guide: VALIDATION_APERCU.md.
• 90 tests passed, 1 ignored. No benchmark on the Acer/RTX 4060 claimed. Player Edit Scene unchanged.

• Input/output handles at the top of audio clips and music; slide inward to lengthen.
• Independent fades saved, limited to the duration of the clip and can be canceled by Ctrl+Z.
• Envelope retained after browsing the preview; applied to export and GPU/CPU/HDR preview.
• Timeline images: same handles and fades as video sequences.
• No changes to Player Edit Scene.
• Validation: 33 targeted tests, including FFmpeg audio envelope, phase after movement and image fades. Windows graphics validation to be performed.