Skip to content

FreeVideo 0.4.9 - #302

Merged
KBRASK merged 19 commits into
mainfrom
release/0.4.9-public
Oct 16, 2026
Merged

KBRASK merged 19 commits into
mainfrom
release/0.4.9-public

Conversation

@KBRASK

@KBRASK KBRASK commented Oct 16, 2026

Copy link
Copy Markdown
Collaborator

0.4.9: Quality levels are reworked: Medium is one pass of 8 steps, and High and Max use step caches to finish much sooner; decoding on Windows reads ahead earlier on slow drives, and the editor handles ranges, splits and text tracks more reliably.

  • Quality levels: Medium is one pass of 8 steps; High and Max run 20 steps with step caches and finish much sooner than plain 20 steps.
  • Decoding on Windows: read-ahead is sized by physical memory and starts earlier on slow drives, so the end of a run waits less for the disk.
  • When Windows runs out of virtual memory, in any system language, FreeVideo now retries the generation automatically in a fresh process and reports a memory shortage if it still fails.
  • The progress estimate counts reference images, so requests with reference images show a closer total time from the start.
  • Editor: dragging a range's left edge to the start of the timeline no longer pushes its right edge later; splitting a sped-up or reversed part keeps the fade-in only on the first half and the fade-out only on the second.
  • Editor: drag-selecting parts is smooth again on timelines with many parts.
  • Editor text: clicking the side Content box right after pressing T keeps the new text track, and dropping text onto a locked text track says the track is locked.
  • Long videos: each part's description can also be edited in a larger window.
  • Chat: when a conversation is full, the notice appears once under the message and its button starts a new chat instead of retrying.
  • Export: the sizes offered in the export dialog are rounded the same way as the saved work.
  • Diagnostic reports record how much of the graphics card other programs were using during each sampling step.
  • Launcher: when the browser cannot be opened automatically, the message says to click Copy address and open it in a browser, in Chinese as well. Canvas margin labels stay readable and aligned. Nodes: width, height, duration and seed have Chinese labels.

KBRASK added 19 commits October 16, 2026 19:02
Continuation is reported unavailable on backends that cannot continue.
Medium is one pass of 8 steps, High uses 20 steps with EasyCache-AV and Max uses 20 steps with the
first-block cache; Mac keeps plain steps. The step cache is carried through requests, identities
and estimates.
Reference video rows are attended by every generated token, so a row costs
several times a generated token and the cost grows with the canvas. The step
estimate counted each row once, which halved the predicted time of requests
with several reference images. Rows now carry a weight that grows with the
canvas up to the measured size, on NVIDIA paths only; requests without
reference rows, audio rows and the Apple fit are unchanged. Adds tests and
registers them in CI.
A Windows "paging file too small" error raised from the safetensors
reader carries only a localized message and no errno, so it was classified
as a code error, shown as a bare worker exit, and left the resident worker
alive with its models still charged to the commit limit.

Recognise error 1455 by code or text in the failure classifier, diagnostics,
launcher failure details and the error panel, and replace the resident
worker after such a failure so the next request starts from a clean
process.
@KBRASK
KBRASK merged commit 3db269a into main Oct 16, 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