A diffusion wardrobe that lets you keep your pose and face

What it does IMAGDressing-v1 is a virtual dressing pipeline built on Stable Diffusion 1.5. You provide a clothing image, and it renders a dressed human, optionally respecting a specific pose, face, or background text prompt. It ships with a Gradio interface and a suite of inference scripts that mix garment generation with IP-Adapter, ControlNet, and T2I-Adapter.

The interesting bit Instead of fine-tuning the whole diffusion model or training a LoRA per outfit, the authors added a dedicated garment UNet that feeds CLIP semantics and VAE textures into a frozen denoising UNet through a hybrid attention layer. This keeps the base model untouched, so off-the-shelf SD plugins still work.

Key highlights

• Accepted at AAAI 2025; releases the IGPair dataset (300k+ clothing-dressed image pairs) and the CAMI affinity metric for benchmarking.
• Modular architecture: garment features plug into a frozen SD backbone, enabling rapid customization “within seconds” without per-garment LoRA training.
• Compatible with the existing ecosystem: IP-Adapter, ControlNet, T2I-Adapter, and AnimateDiff.
• Experimental extras include localized outfit changing via inpainting and cartoon-style generation.
• Third-party ComfyUI and Windows WebUI nodes already exist.

Caveats

• Released checkpoints and downloadable models are limited to non-commercial research use, even though the codebase itself permits broader use.
• Higher-resolution model weights and video dressing are on the TODO list but not yet available.
• Some experimental features rely on additional auxiliary models not included in the main release.

Verdict Researchers and builders working on fashion tech or controllable human generation should take a look. If you need a fully open-weight commercial tool today, the model licensing may give you pause.