Alibaba Cloud
FashionChameleon is a research-oriented AI video project titled FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization. The project page and GitHub repository describe a streaming autoregressive video-generation framework where users can interactively switch garments during generation while preserving motion coherence. The paper highlights teacher-model in-context learning from single reference-garment pairs, streaming distillation with gradient-reweighted distribution matching distillation, and training-free KV-cache rescheduling. Reported performance is 23.8 FPS for 720p generation on a single H200 GPU, described as 30-180x faster than existing customization baselines. The GitHub repository is Apache-2.0 licensed, but checkpoints, pre-training/post-training code and HGC-Bench release are still listed as todo items at verification time.
Editorial verdict
Researchers and e-commerce video teams tracking real-time virtual try-on, garment switching and human-centric video customization methods from Chinese labs.
Avoid treating it as a ready SaaS workflow; wait for checkpoints, runnable code and license checks before production evaluation.
FashionChameleon belongs in AI Video because it is specifically about real-time human-garment video customization, with clear relevance to e-commerce content creation and virtual try-on research.
Research project page and Apache-2.0 GitHub repository; checkpoint and training code releases are still pending
Project page, GitHub repository, arXiv / Hugging Face paper page, HGC-Bench dataset page
Commercial use should follow the current product, API, model license and billing terms.
Review prompt, file, media upload, retention and training-use terms before sensitive workloads.
Use it as a research reference for switching garments during generated human videos while preserving motion coherence.
The paper frames low-latency garment control for e-commerce and content creation.
Teacher-student distillation, gradient-reweighted DMD and KV-cache rescheduling are technical references.
Model names, quotas, release status, regional access and commercial terms can change quickly; recheck official sources before procurement or production use.
kling-ai
seedance-video
hunyuan-open-models
vidu
official · en · verified 2026-05-18
Confirms project title, authors, Alibaba Group and Xiamen University affiliations, method highlights and 23.8 FPS claim.
official · en · verified 2026-05-18
Confirms official PyTorch repository, Apache-2.0 license, release todo list and technical summary.
other · en · verified 2026-05-18
Confirms arXiv identifier, May 2026 paper page, Alibaba submission context and abstract summary.
benchmark · en · verified 2026-05-18
Linked benchmark/dataset destination from the project page.
Last checked: 2026-05-17