README.md

metadata

title: Virtual Try-On ZeroGPU Demo
emoji: 👕
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 5.7.1
python_version: '3.10'
app_file: app.py
pinned: false
license: other
short_description: ZeroGPU xlarge Gradio demo for image-based virtual try-on.

Virtual Try-On ZeroGPU Demo

This Space is a Gradio SDK deployment wrapper for the JoyAI/XVideo virtual try-on inference path.

Inputs

• One multimodal input box for image files and text
• Upload order: first image is the person image; remaining images are garment reference images
• Default prompt buttons: no reference, one reference, two references, and multiple references

Output

• One generated try-on image

ZeroGPU

The inference callback is decorated with:

@spaces.GPU(size="xlarge", duration=180)

Use ZeroGPU xlarge because the model is expected to need at least 60 GB of VRAM.

Required Space Variables

Set these in the Space settings:

MODEL_ID=your-org/your-vton-model

If model files are stored in a Hugging Face Storage Bucket, mount the bucket into the Space and set:

MODEL_LOCAL_DIR=/models

In this mode MODEL_ID is not required. The app loads files directly from the mounted bucket path.

If the model repository is private, add this as a Space Secret:

HF_TOKEN=hf_xxxxxxxxxxxxxxxxx

Optional variables:

MODEL_REVISION=main
JOY_CONFIG_PATH=configs/editing_exp/edit_1024p_demo.py
JOY_TRANSFORMER_PATH=transformer
JOY_DIT_CKPT_TYPE=safetensor
JOY_MODEL_PATH=FireRed-Image-Edit-1.0
JOY_VAE_PATH=FireRed-Image-Edit-1.0/vae/diffusion_pytorch_model.safetensors
QWEN235_CHAT_URL=http://ai-api.jdcloud.com/v1/chat/completions
QWEN235_MODEL=qwen3-vl-235
HF_HOME=/data/.huggingface

For Prompt Enhancer, add one of these as a Space Secret:

QWEN235_API_KEY=your_api_key

or:

OPENAI_API_KEY=your_api_key

Expected Model Repository Layout

This wrapper downloads MODEL_ID with huggingface_hub.snapshot_download. The model repo should contain the JoyAI/XVideo code or a code/ or JoyAI-Image/ folder with the xvideo package. It also needs the config and checkpoints referenced by the variables above.

MODEL_ID repo
├── code/
│   └── xvideo/
├── configs/
│   └── editing_exp/
│       └── edit_1024p_demo.py
├── transformer/
│   ├── model-00001-of-00005.safetensors
│   └── ...
└── FireRed-Image-Edit-1.0/
    ├── processor/
    ├── text_encoder/
    ├── tokenizer/
    └── vae/

If the config/checkpoint files live at different paths, set:

JOY_CONFIG_PATH=relative/or/absolute/config.py
JOY_TRANSFORMER_PATH=relative/or/absolute/transformer_or_step_xxx.pth
JOY_DIT_CKPT_TYPE=safetensor
JOY_MODEL_PATH=relative/or/absolute/FireRed-Image-Edit-1.0
JOY_VAE_PATH=relative/or/absolute/FireRed-Image-Edit-1.0/vae/diffusion_pytorch_model.safetensors
JOY_TEXT_ENCODER_PATH=relative/or/absolute/FireRed-Image-Edit-1.0/text_encoder
JOY_PROCESSOR_PATH=relative/or/absolute/FireRed-Image-Edit-1.0/processor
JOY_TOKENIZER_PATH=relative/or/absolute/FireRed-Image-Edit-1.0/tokenizer

The inference logic follows tryon_infer.py:

processed images = [garment_image, person_image]
prompt = <image><image> + try-on instruction
output = one generated PIL image

Prompt Enhancer is enabled by default in the UI. It sends the primary garment image and person image to the configured Qwen3-VL-235 endpoint, then displays the enhanced prompt used for generation.

If multiple garment images are uploaded, the app currently uses the first garment reference image as the primary garment. It does not pretend to perform multi-reference fusion.

The landing UI includes local visual assets under assets/:

assets/
├── tryon_logo_compact.png
└── showcase/
    ├── case_0.jpg
    ├── case_1.jpg
    ├── case_2.jpg
    └── case_3.jpg

Deployment Steps

1. Create a new Hugging Face Space with SDK Gradio and hardware ZeroGPU.
2. Copy these files into the Space repository.
3. Add any model-specific Python dependencies to requirements.txt.
4. Use one of these model storage options:
   • Model repo: set MODEL_ID as a Space Variable.
   • Storage Bucket: mount the bucket into the Space, for example at /models, then set MODEL_LOCAL_DIR=/models.
5. Set the JoyAI checkpoint/config variables listed above.
6. If the model storage is private, set HF_TOKEN as a Space Secret.
7. Push the Space repository.
8. Open the Space and run a test with one person image and one garment image.

Do not commit large model weights to this Space repository. Keep weights in a Hugging Face Model Repo or a Hugging Face Storage Bucket.

For the complete Storage Bucket deployment flow, including the Davidscut/Tryon-demo-storage mount setup and path variables, see DEPLOYMENT_BUCKET.md.