Skip to content

FreeVideo 0.5.0 - #743

Merged
KBRASK merged 26 commits into
mainfrom
release/0.5.0-public
Nov 11, 2026
Merged

KBRASK merged 26 commits into
mainfrom
release/0.5.0-public

Conversation

@KBRASK

@KBRASK KBRASK commented Nov 11, 2026

Copy link
Copy Markdown
Collaborator

0.5.0: H3 Image arrives: a new Images view for H3 Image and Qwen-Image 2.1, pictures up to 16 MP on graphics cards from 8 GB, and Autopilot pictures follow the installed model. Beat lines join the editor, and launcher failures, settings and cards are clearer.

  • H3 Image: a new image model for NVIDIA graphics cards from 8 GB of video memory, making pictures from 4 MP up to 16 MP. It is installed alone or added to an existing video install, which downloads only the missing files, about 22 GB, and shares the text encoder already installed. One weight file serves every graphics card, and machines with little memory load it in mapped form. Not available on Mac or AMD.
  • Images view: H3 Image and Qwen-Image 2.1 share one prompt box, and finished pictures gather in the Works column. Generated pictures are saved to the asset library and appear under Images in the reference "+" menu. The More menu offers Open folder, Copy prompt and Delete. The empty stage shows sample pictures, and the page has a wider layout, a solid Generate button and a clearer error card. H3 Image does not take reference pictures yet. With the prompt rewrite model installed, the "FreeToken enhancement" switch rewrites H3 Image prompts without adding anything the request did not ask for.
  • Autopilot: picture requests use the installed H3 Image at 4 MP or 16 MP; a request below 4 MP is made at 4 MP in the same shape.
  • Editor: beat lines on the music track, and clips snap to the beats. The feature tour now starts on the clip just imported.
  • Keyframes: still pictures can be placed between the first and last frame, and each is generated one at a time.
  • Audio: with "Use as is" selected and no audio connected, FreeVideo now says so before generating. When audio used as is cannot be generated in parts, the reason is shown and the request stops before it is submitted.
  • Windows: on machines with tight system memory, the second pass no longer stops and restarts for lack of memory. The video is unchanged.
  • Long videos: very long high-resolution requests no longer stop on the first step with a numerical error.
  • Sampling preview: an optional small preview decoder on CUDA, used only when it does not slow generation.
  • Settings: the working files of finished videos can be cleared.
  • Launcher: a failure shows one card with the next step and the buttons that resolve it, and the technical details are folded away. A CUDA initialization failure comes with the step to take, an outdated driver gets a card asking to update the driver, and a failed update records why each source failed.
  • Interface details: checkboxes share one size; the Passes field is as compact as the Reserved VRAM field; media cards no longer show a mode label; editing a character puts the cursor in its name; Autopilot cards show "Custom" for a custom quality and a single progress percentage; and the header and view tabs fit narrow windows without overlap or sideways scrolling.

KBRASK added 26 commits November 3, 2026 08:29
…-driver card

When nvidia-smi cannot answer and the driver predates the CUDA UUID call, the
fallback stopped on a missing function and the launcher showed a generic card.
The fallback now reads the card name and compute capability without the UUID
or the context, so the existing driver-too-old check names the card and asks
for an update. Current drivers are unchanged.
…, stop it before...

Explain why audio used as is cannot be made in parts without two-pass, stop it before submitting, recognise its refusals as known errors, and keep settings in submission reports

Audio longer than 30 seconds used as is is made in several parts, which needs two-pass acceleration. The engine now names that case instead of the generic route message, the studio stops the run with the same sentence before sending it, both refusals are recognised as known errors in reports, and a failed submission report carries the run's settings.
- Browser entry test: follow the shared dropdown wrapper and the empty media mode label; installer test base for the 0.4.38 candidate build
- Failure UI test: the cause text sits in the technical-details box only, the paragraph above it stays empty
At inference the per-frame delta statistics are computed with TF32. For very long, high-resolution requests the rounding can make a few of the matrices indefinite, and the Cholesky factorisation stopped the run. Matrices that factorise keep the existing factor bit for bit; only the ones that fail are rebuilt from their keys and gates in FP64. CUDA only, installed with the inference kernels.
When CUDA is unavailable during GPU detection, read nvidia-smi: a missing tool shows the missing-driver notice, a driver older than 580 the old-driver notice with model and version, and an installed driver whose CUDA still fails (or whose query fails) a new CUDA start notice. The original CUDA error stays in the details.
The passes field in Performance measurement now sits in the same unit row as the reserved VRAM field, 96 px wide and inside its section, instead of stretching past the right edge. The shared-controls browser test opens every section and checks number field width and edge.
Songs with a clear pulse get beat points (confidence >= 0.5). While a clip or
cut point is dragged, thin grey lines mark the beats of the music track and the
dragged edge snaps to them; clip edges, markers and the playhead keep priority.
Beats closer than 32 px are left out. The playhead never snaps to beats.
…he capsule

Checkboxes are 16x16 with a 4px radius everywhere; media card drops the capsule; edited character opens on its name

The shared checkbox rule sat in :where() and lost to the studio input rule and per-panel rules, which stretched the media-card boxes and rounded the LoRA box. Its specificity now outranks them. The media card shows no mode capsule or help line, and the edit dialog returns focus to the name box once the card has loaded.
When a model upgrade fails, the feedback button now sends the full failure text, including the retained step log, instead of only the first line shown on the card.
…g buttons first...

Launcher: one failure card with collapsed technical details, resolving buttons first, single-colour progress bar, no duplicate side card, step line does not repeat the title

The failure card shows the title, the next step and the resolving buttons; the report buttons follow, and the technical details (with the exit code) fold behind one button. A recheck button is named as the text says (Check again / Connect again). The progress bar uses one accent colour, the side card appears only on the launcher page, and the install step line no longer repeats the step title.
… the source manifest check...

Source payload: allow the preview decoder weights, whose absence made the source manifest check reject a Windows build
…der yields in order

The generic header-button style overrode the connection lamp's own pill and its
dot-on-narrow rule; the lamp is now left out of it. Without a running job the
entries beside the view switch had no retreat order and overlapped it on narrower
windows; the idle header now turns the lamp into a dot first, then drops phone,
feedback and tutorial, and uses two rows at 825 px and below, matching the running
bar. Running-bar thresholds move up by the extra width the lamp now takes.
Adds idle-header checks (no overlap at any width, lamp pill kept) to the header test.
…, install alone or add...

Add H3 Image: image view for both image models, one int8 weights file, install alone or add later, 4 to 16 MP

Adds the image view for both image models with a single int8 weights file. The image models can be installed on their own or added later, and images range from 4 to 16 megapixels.
…res on the blank stage...

Image view polish: wider layout without an empty column, sample pictures on the blank stage, always-solid Generate, error card, taller controls, prompt focus ring
@KBRASK
KBRASK merged commit 7c808ca into main Nov 11, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant