diffUIsion is a lightweight FastAPI + Diffusers web UI for image-generation models (FLUX family by default). It can optionally use a bundled local language-model to “enhance” your prompt before the image is rendered (no external API needed).
- One-page Bootstrap UI with live token counter
- Optional prompt enhancement via a local model
- Model/height/width/etc. taken from
models.toml– add your own entries - Generation runs in an isolated subprocess → cancel button frees GPU RAM instantly
- Fully reproducible (seed input)
- Pure Python, works on CPU, CUDA, MPS & XPU
- Clone & enter
git clone https://github.com/MattKobayashi/diffuision.git
cd diffuision- Add secrets
echo "<YOUR_HF_TOKEN>" > secrets/hf_token- Run
uv run web.pyOpen http://localhost:8000 in your browser.
| File / Environment Variable | Purpose |
|---|---|
secrets/hf_token |
HuggingFace token for model downloads |
Each [[models]] table describes overrides for height/width, number of
steps, guidance scale, etc., keyed by the HuggingFace repo id (name).
Pressing Cancel calls /cancel, which terminates the child
generation process. GPU/CPU memory is released immediately and the UI
is ready for another prompt.
├── web.py ← main FastAPI application
├── templates/ ← Jinja2 templates (index & result)
├── generated/ ← saved images (auto-created)
├── secrets/ ← API tokens (ignored by git)
└── models.toml ← per-model defaults