huggingface-projects/diffusiongemma-codegen

Gemma Diffusion · Live Website Builder

A live, side-by-side visualization of block-diffusion code generation. Describe a website and the model writes a single self-contained HTML document by denoising a canvas of random tokens — every token position updates at once each step.

The left pane shows the raw HTML canvas diffusing token-by-token (the signature look of text diffusion); the right pane renders the page live. Type a follow-up to tweak in place: the previous page seeds the diffusion's starting canvas (via the model's native canvas_ids API), so the model edits the existing page instead of regenerating from scratch. Changed lines are highlighted, and the preview keeps your scroll position across re-renders.

Stack

• —Model: google/diffusiongemma-26B-A4B-it (DiffusionGemmaForBlockDiffusion).
• —Backend: `gradio.Server` — a FastAPI subclass that provides Gradio's queue + SSE streaming under a custom, hand-written HTML/CSS/JS frontend (index.html). The single streaming endpoint /generate yields one JSON frame per denoising step.
• —Hardware: ZeroGPU (xlarge) — the 26B checkpoint needs the full backing card.

A custom transformers wheel providing the DiffusionGemma architecture is bundled in this repo and installed at runtime by app.py (Spaces installs requirements.txt before the repo files are copied in, so a local-path wheel can't be referenced there).

Configuration

• —HF_TOKEN (secret) — read access to the private model repo.
• —GRADIO_SSR_MODE=false (variable) — required so the custom / route serves index.html instead of Gradio's SSR shell.
• —GDIFF_MODEL_PATH (optional) — override the model repo id.
• —GDIFF_GPU_SIZE (optional) — ZeroGPU slice, defaults to xlarge.