Qwen Image Edit — Rapid AIO Web UI

A web interface (in the style of the official Qwen/Qwen-Image-Edit-2511 Space) for Phr00t/Qwen-Image-Edit-Rapid-AIO — a Lightning-merged FP8 build of Qwen-Image-Edit-2511 that edits images in ~4 inference steps with CFG 1.0. Default checkpoint: v23 NSFW.

Only the transformer is taken from the AIO checkpoint (loaded with QwenImageTransformer2DModel.from_single_file, FP8 storage with bf16 compute via layerwise casting); the text encoder / VAE / processor come from the base Qwen/Qwen-Image-Edit-2511 repo. No ComfyUI needed. GGUF checkpoints (e.g. Phil2Sat/Qwen-Image-Edit-Rapid-AIO-GGUF) are also supported via MODEL_URL.

Architecture — two independent deployments

• Modal (complete: UI + inference): a GPU backend (default: one L40S, scales to zero when idle) serving the Gradio web UI and a JSON API at https://<workspace>--qwen-image-edit-rapid-web.modal.run.
• Hugging Face Space (complete: UI + inference): the same Gradio app running the model directly on the Space via ZeroGPU (requires PRO). The Space can instead be deployed as a thin client for the Modal backend (SPACE_MODE=remote).
• GitHub Actions deploys both — everything can be done from a phone browser. There is also scripts/deploy_via_hf.py for environments where the Modal CLI cannot connect (it sets up a private "deployer" Space that runs modal deploy from Hugging Face infrastructure).

Deploy (works entirely from a phone)

1. Get tokens

• Modal: sign up at modal.com, then go to Settings → API Tokens → New Token. Copy the token ID (ak-…) and secret (as-…).
• Hugging Face: Settings → Access Tokens → create a token with Write access.

2. Add them to this GitHub repo

Settings → Secrets and variables → Actions → New repository secret:

When MODAL_AUTH_TOKEN is set, the API requires an Authorization: Bearer <token> header, and the Modal-hosted web UI asks for a login: username user, password = the token. The Space authenticates automatically via its MODAL_AUTH_TOKEN secret.

3. Run the workflow

Actions tab → Deploy → Run workflow (it also runs automatically on every push to main). The job summary shows both URLs when done:

• Modal web UI + API: https://<workspace>--qwen-image-edit-rapid-web.modal.run
• Space: https://huggingface.co/spaces/<you>/qwen-image-edit-rapid

The first deploy includes a warm-up run that downloads ~35 GB of weights into a Modal volume (can take 15–30 min). After that, cold starts are much faster and the container scales to zero after 5 idle minutes, so you only pay while generating.

Optional repository variables

Settings → Secrets and variables → Actions → Variables:

API

curl -X POST "https://<workspace>--qwen-image-edit-rapid-web.modal.run/v1/edit" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $MODAL_AUTH_TOKEN" \
  -d '{
        "prompt": "make the sky a dramatic sunset",
        "images": ["<base64-encoded png or jpeg>"],
        "num_inference_steps": 4,
        "true_cfg_scale": 1.0
      }'

Response: {"images": ["<base64 png>", …], "seed": 123}. Up to 3 input images and 4 outputs per request. Omit seed for a random one.

Run locally (GPU with ~40 GB VRAM, or less with LOW_VRAM=1)

pip install -r requirements.txt
python app.py            # local inference
# or use a remote backend:
MODAL_ENDPOINT_URL=https://…modal.run python app.py

Notes

• Default model file: v23/Qwen-Rapid-AIO-NSFW-v23.safetensors (28.4 GB, FP8). Phr00t's notes: v19 is best for edit consistency, v23 for prompt adherence; each version has SFW and NSFW variants. Switch with the MODEL_URL variable.
• Rapid settings: 4–8 steps, True CFG 1.0, no negative prompt needed.