FashionChameleon: A Real-Time AI Framework for Interactive Garment Customization in Video

Summary

This paper presents FashionChameleon, a real-time and interactive framework for human-garment video customization. It addresses the challenge of enabling users to interactively switch garments during video generation while preserving motion coherence, using only single-garment video data (not multi-garment data). The framework introduces three key techniques: (1) a Teacher Model with In-Context Learning trained on single reference-garment pairs, (2) Streaming Distillation with In-Context Learning for consistency and efficiency, and (3) Training-Free KV Cache Rescheduling for interactive multi-garment switching. FashionChameleon achieves real-time generation at 23.8 FPS on a single GPU, making it 30-180x faster than existing baselines, with potential applications in e-commerce and content creation.