Research map / Image generation / Editing and style transfer
IMAGDressing-v1: Customizable Virtual Dressing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Defines the virtual dressing (VD) task for merchants, proposes IMAGDressing-v1 (garment UNet + hybrid attention in a frozen SD1.5 denoising UNet), a CAMI metric, and the 324,857-pair IGPair dataset.
Problem
Existing virtual try-on methods offer limited control over garment presentation, faces, poses, and scenes, so merchants cannot flexibly showcase garments in generated human images.
Approach
A trainable garment UNet (initialized from SD v1.5) extracts semantic CLIP features and VAE texture features and injects them into a frozen denoising UNet via a hybrid attention module (frozen self-attention plus trainable garment cross-attention), keeping text-driven scene control and making it pluggable with ControlNet and IP-Adapter.
Key results
- IMAGDressing-v1 reaches ImageReward -0.095, MP-LPIPS 0.1466, CAMI-U 1.753 and CAMI-S 2.719, the best among compared methods (MagicClothing: -0.164, 0.1499, 1.655, 2.692).
- Ablation: adding the image encoder branch and hybrid attention raises CAMI-U from 1.575 (base) to 1.753 and lowers MP-LPIPS from 0.1537 to 0.1466.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace MagicClothing with IMAGDressing-v1 | Virtual dressing methods · Virtual dressing | ImageReward: -0.164 -> -0.095 | Helped |
| replace MagicClothing with IMAGDressing-v1 | Virtual dressing methods · Virtual dressing | MP-LPIPS: 0.1499 -> 0.1466 | Helped |
| replace MagicClothing with IMAGDressing-v1 | Virtual dressing methods · CAMI-U | CAMI-U: 1.655 -> 1.753 | Helped |
| replace MagicClothing with IMAGDressing-v1 | Virtual dressing methods · CAMI-S | CAMI-S: 2.692 -> 2.719 | Helped |
| add image encoder branch (IEB) (ablation) | IMAGDressing-v1 base · Virtual dressing | ImageReward: -0.245 -> -0.178 | Helped |
| add image encoder branch (IEB) (ablation) | IMAGDressing-v1 base · Virtual dressing | MP-LPIPS: 0.1537 -> 0.1504 | Helped |
| add image encoder branch (IEB) (ablation) | IMAGDressing-v1 base · CAMI-U | CAMI-U: 1.575 -> 1.637 | Helped |
| add image encoder branch (IEB) (ablation) | IMAGDressing-v1 base · CAMI-S | CAMI-S: 2.578 -> 2.625 | Helped |
| add hybrid attention (HA) to the IEB variant (ablation) | IMAGDressing-v1 · Virtual dressing | ImageReward: -0.178 -> -0.095 | Helped |
| add hybrid attention (HA) to the IEB variant (ablation) | IMAGDressing-v1 · Virtual dressing | MP-LPIPS: 0.1504 -> 0.1466 | Helped |
| add hybrid attention (HA) to the IEB variant (ablation) | IMAGDressing-v1 · CAMI-U | CAMI-U: 1.637 -> 1.753 | Helped |
| add hybrid attention (HA) to the IEB variant (ablation) | IMAGDressing-v1 · CAMI-S | CAMI-S: 2.625 -> 2.719 | Helped |
Limitations
- Comparisons are limited to four baselines, and CAMI-S cannot be computed for several of them.
- Garment strength lambda must be tuned, trading off text adherence against garment fidelity.
- The CAMI sub-score details are deferred to the supplementary material, so the metric is not fully specified in the main text.
Experimental setup
Datasets: IGPair
Benchmarks: IGPair
Metrics: ImageReward MP-LPIPS CAMI-U CAMI-S
Baselines: BLIP-Diffusion Versatile Diffusion IP-Adapter MagicClothing
Builds on: Stable Diffusion v1.5 ControlNet IP-Adapter classifier-free guidance
Closest papers by meaning
- M&M VTO: Multi-Garment Virtual Try-On and Editing (CVPR 2024 · similarity 0.90)
- AnyDressing: Customizable Multi-Garment Virtual Dressing via Latent Diffusion Models (CVPR 2025 · similarity 0.90)
- Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling (CVPR 2025 · similarity 0.89)
- PICTURE: PhotorealistIC virtual Try-on from UnconstRained dEsigns (CVPR 2024 · similarity 0.89)
- MV-VTON: Multi-View Virtual Try-On with Diffusion Models (AAAI 2025 · similarity 0.89)
- OOTDiffusion: Outfitting Fusion Based Latent Diffusion for Controllable Virtual Try-On (AAAI 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.