Repository navigation
FreeVideo 0.4.34 - #663
Merged
Merged
FreeVideo 0.4.34#663
Conversation
…empo change closes a line... Voice change: dubbed lines keep the original pace and loudness: the tempo change closes a line first and each new line comes back as loud as the one it replaces
…shots, length from the... Drama: numbered cuts in one shot, six reference sections for picture shots, length from the capability table, speaker numbers by order in the video, continuity table in every shot, defined and capitalised names, blank notes
…ed canvas never shows black
…port capability limits as... Agent skill: submit storyboards as one batch of independent shots, report capability limits as given, resume retryable failures once, nearest legal size on invalid_request, trim linked sound with the clip
…locking, continuity... Agent skill: short-drama craft (story outline, 15-20 s generations, blocking, continuity, casting, prompt rules, five pre-generation checks) and the official prompt details: speaker numbers by order in the video, summary task type, style before Shot 1, cut-crossing lines, voice citation
…judged by its limit, also... Drama: a long shot waits for this computer's capability table and is judged by its limit, also in the preview and reshoots
Shots that share a voice reference with different prompts no longer reload the audio VAE. The normalized latent is kept in the input cache under the samples, their shape, the 32 kHz rate and the audio VAE identity; a clip used as is trims after the lookup, so a hit and a miss give identical bytes.
…s the recognised text, and... Voice change: lines are numbered by who says them, known text replaces the recognised text, and a text that cannot fit its room is named before anything is made
The clip's sound track is encoded like a plain audio reference and no picture is encoded, so a lone voice-only clip uses the audio task and is tagged <Audio N> in the Agent guide. A clip without sound gets a plain error.
…ill contact sheet The sampler's coarse preview now carries up to sixteen evenly spaced latent frames (was four) and the page shows one frame in place, playing forward at 4-8 frames a second over the video's length. The frame takes only the height left in the stage after the percentage, step and times; reduced motion keeps the last frame still. The phone's large preview plays too; its list picture stays still. Sampling is unchanged: the preview only reads tensors the sampler already produced.
…inese and English), a... Take QA: script lines are checked with the real speech recogniser (Chinese and English), a single odd frame at the start or end is noted, a picture that stops before the end is a problem
…h clip's sound, remove... Agent: finish the sound of an editor project in one job: separate each clip's sound, remove sustained music, lay one continuous score, even out loudness and fill silent gaps with room tone
…rom the same pictures... Videos with the same prompt fold together only within two hours and from the same pictures, clips and audio The same prompt text linked every video of that day into one task, however many hours apart and whatever reference pictures, clips or audio each was made from. A prompt now links videos made within two hours of the one before it, across midnight too, and from identical inputs; previews and upscales, continuations and agent jobs link as before.
… the film-wide row for that... A film script's continuity row can be set in a single shot, replacing the film-wide row for that shot only, so a state that changes during the film is written where it holds
…o report a problem and... Launcher: the step that asks for a ComfyUI restart no longer offers to report a problem and names the button it needs, "Connect again"
…onger offers to report a... Launcher: the wrong-folder card points to installing ComfyUI and no longer offers to report a problem When no ComfyUI is found in or near the chosen folder, the card now also says to choose "Install ComfyUI" if it is not installed yet. A wrong folder is something to correct, not a fault to report, so the card and the error box in Settings drop "Report a problem" like the taken-port card, and the card keeps Copy details, Show details and Export report.
… GPU memory The refusal's 'Fits now' line only tried 0.75 and 0.5 megapixels, so a 2 MP request on a 6 GB card was told to use 1184x672 although a larger size was admitted by the same planner. Try 1.5, 1.0, 0.75 and 0.5 megapixels in that order. Only the refusal text changes; placement and budgets are untouched.
…load for the dubbing and... Voice lines: offer the dubbing model download in the editor, one download for the dubbing and speech models, plain recognition failures The voice change panel offers the dubbing model download when this machine can use it, with one progress for the dubbing pack and the three speech groups. A recognition that cannot run says why in the panel: a model that went missing brings the download back, an unusable one and other failures give one sentence with a report button, an unsupported language gives its own sentence. A cancelled recognition leaves no panel behind.
…ce, and the download list... Launcher: a setup check that ended without output says so in a sentence, and the download list actions stack wherever one row cannot hold every label The check's summary line stays in the copied details and the report but no longer shows on the failure card or under the download list. The download list's buttons form one full-width column at phone width or wherever the row would wrap a label, and keep their own width in the row otherwise. The layout test holds the rule under any font metrics, including a machine that draws every letter one em wide.
…eo encoding reports each... Reference media: the estimate counts the prepared references, and video encoding reports each finished chunk A reference or guide video enlarges every sampling step and is the longest silent stage of a request on a modest GPU. The estimate started from the request alone, so it ignored the reference rows and gave the encoding stage a fixed few seconds, and the display did not move until the stage ended. The worker now reports the prepared references' size right after normalization, and each temporal chunk of the video encoder reports its progress with the elapsed time; the tracker rescales the stage under way from that rate. The encoded latents are unchanged.
Normalized reference frames (over 1 GB per reference video) were kept in every creation folder, including cancelled and failed ones that have no MP4. The generator now deletes them once the encoder is done, the release also covers them in existing folders, and stopped requests are released at the request end and at the next start. A frame preview left by a stopped request no longer makes its folder count as foreign. A frame file that Windows still holds open no longer fails the request.
…ading, quiet on a first... What's-new pages for 0.5: animated track editor shown once after upgrading, quiet on a first launch, take-me-there tours for the three sound chapters
The page changes voices through the project voice routes and reads whether it can from the agent root route, but only the token was accepted there, so on a real server the voice entry never appeared. These routes now accept the page session like the sound-separation routes, and a test scans the page's own agent calls so a route it calls cannot lack that access again.
…oice how-tos, first-time... Tutorial: sound chapters, separate-tracks / replace-music / replace-voice how-tos, first-time bubble on Separate tracks, and the Separate and replace sound card
…ith segment buttons that... Tour cards: add sound to a video and generate a picture from sound, with segment buttons that play each segment with sound
…sed start reads as a quiet wait The Run placement test button stayed enabled during the installation check and during installation. The engine then refused the start with a wait message, which the panel showed as an error with a report button. The button is now disabled whenever a setup task is busy and enabled again when the polled state is idle. The refusal that names the running task has an entry in the setup messages, so it is shown as a plain sentence (translated) without the report button, in the muted style of the starting message. After a refusal the installation section opens and the state is polled so the progress it points to is visible, and the sentence is cleared when the task ends.
…ad as a picture, kept in... Editor: a button next to the preview saves the frame under the playhead as a picture, kept in the image history
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
0.4.34: Agent can finish the sound of a whole editor project in one job, a line of a clip can be respoken in the original voice, and the Edit tab can save the current frame as a picture.