Video garment swapping is usually an offline, per-clip task. FashionChameleon makes it interactive and real-time: if a person in a video wears a red hoodie, you can swap it to a blue one — not just in a single image, but across every frame — with the blue garment's texture deforming naturally as the person walks or turns their head.
The system runs during video generation at 23.8 FPS on a single GPU, and the paper is available as arXiv:2605.15824.
Three Core Techniques
1. Single-garment, single-person training with in-context learning. The model is trained using only paired data of one garment on one person. In-context learning lets the model implicitly maintain garment consistency across frames despite this limited supervision.
2. Streaming distillation + teacher-forcing fine-tuning. Streaming distillation compresses the model toward real-time inference, while teacher-forcing fine-tuning over the in-context learning setup preserves consistency quality after acceleration.
3. Training-free KV-cache rescheduling. A mechanism applied at inference time that enables interactive garment swapping without any additional training.
The authors report a 30-180x speedup over existing methods.
Open Questions Raised in the Post
- To what extent does "motion consistency" during garment swapping depend on the diversity of poses in the reference video? If the reference person only appears in a standing pose, can garment texture stay consistent during generated head-turning motion?
- The paper uses in-context learning to mitigate this, but does not state the performance boundary across different motion amplitudes.
References
1. Song, Q., et al. (2026). *FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization*. arXiv:2605.15824 [cs.CV]. 2. Esser, P., et al. (2023). *Structure and Content-Guided Video Synthesis with Diffusion Models*. 3. Ma, Y., et al. (2024). *MagicAnimate: Temporally Consistent Human Image Animation*.