Skip to content

Add optional FreeToken prompt enhancement - #75

Merged
KBRASK merged 3 commits into
mainfrom
feat/freetoken-prompt-enhancement
Oct 6, 2026
Merged

KBRASK merged 3 commits into
mainfrom
feat/freetoken-prompt-enhancement

Conversation

@KBRASK

@KBRASK KBRASK commented Oct 6, 2026 •

Copy link
Copy Markdown
Collaborator

Change / 改动

Add optional FreeToken prompt enhancement to the creation panel. Users can rewrite their text and reference images for MiniMax H3, keep both drafts editable, and choose the original or enhanced prompt for generation.

  • Off by default; first use asks before downloading the local Qwen3-VL-4B-Instruct model (about 8.3 GiB).
  • Compact controls inside the prompt editor, with original/enhanced selection and a rewrite action.
  • Adapts FreeToken's portable inference optimizations with Apache 2.0 attribution. The isolated rewrite process releases its resources before video generation, including after cancellation or failure.
  • Adds timing, resource use and structured failure details to the existing local video report, excluding prompt contents, reference images, private paths and raw error text.

Validation / 验证

  • Public installation checks, all browser-module syntax checks and the public entry/prompt-guide checks passed.
  • Built both sdist and wheel; verified the optional model catalog, code and license are packaged, with no model weights included.
  • Development verification: 53 targeted Python checks, relevant JavaScript checks and Chinese/English Chromium interactions passed. Real PRO 6000 tests covered completion, cancellation, reference images and CPU offload, with no remaining rewrite GPU process.
  • Fixed-length comparison measured about 6% higher decoding throughput, but no end-to-end improvement for the short request (8.64 s stock versus 8.75 s adapted).

Review updates

  • Cancel stays available for the whole model download and rewrite; the download shows GiB with a progress meter, preparing and rewriting show an indeterminate meter, and errors are shown in the danger colour.
  • The download consent names the model (Qwen3-VL-4B).
  • Release notes: 0.2.5 now covers prompt enhancement together with the installer retries and download cache cleanup (Fix installer retries and add download cache cleanup (v0.2.5) #73).

Native Windows validation follows on the nightly build before a stable release. This initial enhancement supports text and image references, not video or audio references.

KBRASK and others added 3 commits October 6, 2026 12:07
…e 0.2.5 notes

Cancel was hidden for the whole download and rewrite because the consent closed
without re-rendering the bar. The model download now shows GiB and a meter,
preparing and rewriting show an indeterminate meter, errors use the danger
colour, and the consent names the model it downloads. The 0.2.5 release notes
cover prompt enhancement together with the installer retries and cleanup.
@KBRASK
KBRASK marked this pull request as ready for review October 6, 2026 12:41
@KBRASK
KBRASK merged commit 5f26105 into main Oct 6, 2026
4 checks passed
@KBRASK
KBRASK deleted the feat/freetoken-prompt-enhancement branch October 6, 2026 12:41
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