FastH3-Omni: Gradio 6 reference composer UI

#1
by multimodalart HF Staff - opened

Frontend-only rework, backend (fasth3_engine.py / fasth3_ref.py) unchanged apart from generate taking ordered (kind, path) references.

  • Composer: a custom Gradio 6 gr.HTML component (composer.html / .css / .js). Drop images, videos or audio; each reference gets its <Picture n> / <Video n> / <Audio n> tag on upload, drag to reorder (tags renumber), click a tag to insert it.
  • Builder / Write: the builder assembles the six-section H3 prompt from a description per reference plus scene, soundscape, music and style, with a live preview; Write is free text with tag chips and a lint (missing / unknown / removed tags).
  • Upsample: rewrites through hugging-apps/minimax-h3-prompt-rewriter (visitor's ZeroGPU quota, token=False) and loads the result into Write.
  • Gradio 6.30: FastVideo comes from a wheel of the same commit 376e445f with its unused gradio==5.32.0 pin relaxed (multimodalart/zerogpu-blackwell-wheels). SSR off so the component's listeners survive hydration.
  • Branding: renamed to FastH3-Omni, FastVideo logo blue #346cff, black in dark mode. README models: lists every model of the FastH3-Omni collection; the header links the 8-step V1 model card and the collection.
  • Reference paths coming back from the component are only accepted from Gradio's upload folder or examples/.

Verified on a private staging copy: builds and boots on Gradio 6.30, composer uploads / tags / examples work on the live page.

multimodalart changed pull request title from UI from multimodalart/FastH3-Omni: ordered image/audio/video references, upsample, FastVideo colors to FastH3-Omni: Gradio 6 reference composer UI
multimodalart changed pull request status to merged

Sign up or log in to comment