Research map / Video / Video generation
FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization
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.
Proposes FashionChameleon, a real-time autoregressive framework for interactive human-garment video customization that switches garments mid-generation using single-garment training, in-context distillation, and training-free KV cache rescheduling.
Problem
Existing multi-reference video customization methods are slow and non-interactive, and multi-garment switching videos are scarce for training, so garment switching during generation while keeping motion coherent is hard.
Approach
Train a bidirectional teacher on single reference-garment pairs with in-context conditioning and mismatched garments, distill it into a few-step autoregressive student with in-context teacher forcing and entropy-style frame reweighted DMD, then switch garments at inference via KV cache refresh, historical KV withdrawal, and reference KV disentanglement.
Key results
- FashionChameleon (5B) reaches 23.8 FPS on an H200, versus 0.13-0.77 FPS for the 14B and 1.3B baselines, per the paper's efficiency comparison.
- Achieves the best HGC (4.6833), LGC (3.9250), NTP (4.7625), Smoothness (0.9969) and VQ (0.7483) among methods in the main table on 240-sample HGC-Bench.
- Ablation: GR-DMD with tau=0.2 improves Amplitude from 0.8026 to 0.8395 and Smoothness from 0.9932 to 0.9975 over naive DMD on 165-frame videos.
Surprising findings
- Garment KV refresh alone fails to change the garment because the model attends more to historical KV than to conditioning KV, motivating historical KV withdrawal.
Limitations
- Limited garment categories and variations in training data restricts generalization to complex scenarios.
- Struggles with complex human motions and camera movements, attributed to the Wan backbone.
- Experiments use a self-built 240-sample benchmark judged largely by Gemini-3.0 and a user study.
Experimental setup
Datasets: internet-collected videos curated into 62K reference-garment-video triplets HGC-Bench
Benchmarks: HGC-Bench OpenS2V-Nexus metrics VBench VQ
Metrics: Cur (ArcFace ID similarity) GME Amplitude Smoothness VQ (MUSIQ) HGC LGC NTP FPS
Baselines: VACE Kaleido MAGREF SkyReels-A2 Phantom Qwen-Image-Edit + Wan2.2-TI2V-5B
Builds on: Self Forcing CausVid Distribution Matching Distillation Wan2.2-TI2V-5B Causal Forcing
Closest papers by meaning
- IMAGDressing-v1: Customizable Virtual Dressing (AAAI 2025 · similarity 0.89)
- Controllable Human Image Generation with Personalized Multi-Garments (CVPR 2025 · similarity 0.88)
- Pursuing Temporal-Consistent Video Virtual Try-On via Dynamic Pose Interaction (CVPR 2025 · similarity 0.88)
- AnyDressing: Customizable Multi-Garment Virtual Dressing via Latent Diffusion Models (CVPR 2025 · similarity 0.88)
- SwiftTry: Fast and Consistent Video Virtual Try-On with Diffusion Models (AAAI 2025 · similarity 0.87)
- FashionTailor: Controllable Clothing Editing for Human Images with Appearance Preserving (AAAI 2025 · similarity 0.87)
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.