[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-78824":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":17,"stars7d":18,"stars30d":19,"stars90d":16,"forks30d":16,"starsTrendScore":20,"compositeScore":21,"rankGlobal":10,"rankLanguage":10,"license":22,"archived":23,"fork":23,"defaultBranch":24,"hasWiki":25,"hasPages":23,"topics":26,"createdAt":10,"pushedAt":10,"updatedAt":33,"readmeContent":34,"aiSummary":35,"trendingCount":16,"starSnapshotCount":16,"syncStatus":15,"lastSyncTime":36,"discoverSource":37},78824,"FashionChameleon","QuanjianSong\u002FFashionChameleon","QuanjianSong","Official Pytorch Code of the Paper \"FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization\"","https:\u002F\u002Fquanjiansong.github.io\u002Fprojects\u002FFashionChameleon\u002F",null,"Python",245,12,7,2,0,3,5,98,11,61.64,"Apache License 2.0",false,"main",true,[27,28,29,30,31,32],"fashion","garment-switch","interactive","real-time","streaming","video-customization","2026-06-12 04:01:24","\u003Cdiv align=\"center\">\n\u003Ch1>\n\u003Cimg src=\"assets\u002Ffashionchameleon.png\" width=\"40\" style=\"vertical-align: middle;\" \u002F>\nFashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization\n\u003C\u002Fh1>\n\n\u003Cp align=\"center\">\n    \u003Cspan>\n        \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fpdf\u002F2605.15824\" target=\"_blank\"> \n        \u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F2605.15824-FashionChameleon-red' alt='Paper PDF'>\u003C\u002Fa>      \n    \u003C\u002Fspan>\n    \u003Cspan> \n        \u003Ca href='https:\u002F\u002Fquanjiansong.github.io\u002Fprojects\u002FFashionChameleon' target=\"_blank\">\n        \u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProject_Page-FashionChameleon-green' alt='Project Page'>\u003C\u002Fa>      \n    \u003C\u002Fspan>\n    \u003Cbr>\n    \u003Cspan> \n        \u003Ca href='https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2605.15824' target=\"_blank\"> \n        \u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FHugging_Face-FashionChameleon-blue' alt='Hugging Face'>\u003C\u002Fa>     \n    \u003C\u002Fspan>\n    \u003Cspan> \n        \u003Ca href='https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002FQuanjianSong\u002FHGC-Bench' target=\"_blank\"> \n        \u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FHugging_Face-HGC--Bench-yellow' alt='HGC-Bench'>\u003C\u002Fa>     \n    \u003C\u002Fspan>\n\u003C\u002Fp>\n\n\u003Cbr\u002F>\n\n\u003Cdiv align=\"center\">\n\u003Cb>TL;DR:\u003C\u002Fb>\u003Cbr\u002F>\nWe propose \u003Cspan className=\"text-white font-medium\">FashionChameleon\u003C\u002Fspan>, a real-time and interactive framework for human-garment customization in streaming autoregressive video generation.  \nIt achieves real-time generation at 23.8 FPS on a single GPU.\n\u003C\u002Fdiv>\n\u003Cimg src=\"assets\u002Fteaser.png\" style=\"width:100%; height:100%;\"\u002F>\n\n\n\u003C\u002Fdiv>\n\n\n## 📅 Todo\n- [ ] Release the checkpoint.\n- [ ] Release the training-free kv cache rescheduling for interactive inference.\n- [ ] Release the code (Wan2.2-TI2V-5B) for gradient-reweighted dmd and the corresponding inference.\n- [ ] Release the code (Wan2.2-TI2V-5B) for in-context teacher forcing and the corresponding inference.\n- [x] 🔥 Release the code (Wan2.2-TI2V-5B) for in-context sft and the corresponding inference.\n- [x] 🔥 Release the HGC-Bench.\n- [x] 🔥 Release the \u003Ca href=\"https:\u002F\u002Fquanjiansong.github.io\u002Fprojects\u002FFashionChameleon\u002F\" target=\"_blank\">Project Page\u003C\u002Fa>.\n- [x] 🔥 Release the \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fpdf\u002F2605.15824\" target=\"_blank\">Technical Report\u003C\u002Fa>.\n\n\n## ✨ Highlight\n> **1. Interactive Customization.** We train a single-garment switching teacher using tailored I2V priors and mismatched reference–garment pairs. During generation, we introduce KV-cache rescheduling to enable interactive multi-garment customization without requiring video data containing multi-garment switching.\n\n> **2. Gradient-Reweighted DMD.** Traditional self-forcing treats all self-rolled frames equally during DMD backpropagation. However, later frames typically suffer from larger quality degradation and thus require stronger gradient supervision. We dynamically reweight DMD gradients during self-rolling using a reward model to improve extrapolation consistency.\n\n> **3. Real-Time Generation.** Through streaming distillation with in-context learning, FashionChameleon achieves 23.8 FPS for 720p generation on a single H200 GPU, 30–180× faster than existing customization methods.\n\n\u003Cimg src=\"assets\u002Fintro.png\" style=\"width:100%; height:100%;\"\u002F>\n\n\n## 🎬 Overview\n***FashionChameleon*** is built upon [Wan2.2-TI2V-5B](https:\u002F\u002Fhuggingface.co\u002FWan-AI\u002FWan2.2-TI2V-5B), featuring: **(i)** Teacher Model with In-Context Learning, **(ii)** Streaming Distillation with In-Context Learning, and **(iii)** Training-Free KV Cache Rescheduling.\n\u003Cimg src=\"assets\u002Foverall_framework.png\" style=\"width:100%; height:100%;\"\u002F>\n\n\n## 🔧 Step0. Setup\n### Prepare Environment\n```\ngit clone https:\u002F\u002Fgithub.com\u002FQuanjianSong\u002FFashionChameleon.git\ncd FashionChameleon\n# Installation with the requirement.txt\nconda create -n FashionChameleon python=3.10\nconda activate FashionChameleon\npip install -r requirements.txt\n```\n\n### Download Backbone\nOur FashionChameleon is built upon [Wan2.2-TI2V-5B](https:\u002F\u002Fhuggingface.co\u002FWan-AI\u002FWan2.2-TI2V-5B). You should first download the backbone weight by running:\n```\nhuggingface-cli download Wan-AI\u002FWan2.2-TI2V-5B --local-dir-use-symlinks False --local-dir wan_models\u002FWan2.2-TI2V-5B\n```\n\n## 🚀 Step1. In-Context SFT  \n### Start Training\nYou can run the following command to start training:\n```\nCUDA_VISIBLE_DEVICES=4,5,6,7 torchrun --nproc_per_node 4 --master_port=8989 train.py \\\n    --config_path configs\u002Fsft_wan22_ic.yaml \\\n    --save_dir outputs\u002Fsft_wan22_ic\n```\nor simply run:\n```bash\nbash scripts\u002Ftrain\u002Fsft_wan22_ic.sh\n```\nAll training configurations are recorded in `configs\u002Fsft_wan22_ic.yaml`, which can be freely modified according to your needs.\nNote that our training framework supports both **variable-resolution bucketing strategies** and **gradient accumulation**, you only need to adjust the corresponding `ASPECT_RATIO` and `grad_accum_steps` parameters.\nOur FashionChameleon keep a fixed training resolution of 1280 × 704 throughout training.\n\n### Start Inference\nYou can run the following command to start training:\n```\nCUDA_VISIBLE_DEVICES=1 python infer_ic.py --config_path configs\u002Fsft_wan22_ic.yaml \\\n    --seed 42 \\\n    --h 1280 \\\n    --w 704 \\\n    --num_frames 81 \\\n    --output_path samples\u002Fsft_wan22_ic\u002F \\\n    --checkpoint XXX\n```\nor simple run:\n```bash\nbash scripts\u002Finfer\u002Finfer_wan22_ic.sh\n```\nThe checkpoint represents the weights after SFT training.\nOur inference code by default processes data in the format of HGC-Bench. You can first download the test dataset from [Hugging Face](https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002FQuanjianSong\u002FHGC-Bench).\n\n\u003C!-- ## 🔧 Step2-In-Context Teacher Forcing\nXXX\n\n## 🔧 Step3-Gradient-Reweighted DMD\nXXX -->\n\n## 🌈 Comparison\n\u003Cimg src=\"assets\u002Fcomparison.png\" style=\"width:100%; height:100%;\"\u002F>\n\n\n\n## 🌊 Application\n\u003Cimg src=\"assets\u002Fapplication.png\" style=\"width:100%; height:100%;\"\u002F>\n\n\n## 🤝 Acknowledgements\nThis codebase borrows from [Wan2.2](https:\u002F\u002Fgithub.com\u002FWan-Video\u002FWan2.2) and [Self-Forcing](https:\u002F\u002Fgithub.com\u002Fguandeh17\u002Fself-forcing). Many thanks to them for open-source contributions.\n\n\n## 🎓 Bibtex\n🤗 If you find this code helpful for your research, please cite:\n```\n@article{song2026fashionchameleon,\n  title={FashionChameleon: Towards Real-Time and Interactive Human-Garment Video Customization},\n  author={Song, Quanjian and Shen, Yefeng and Chen, Mengting and Sun, Hao and Lan, Jinsong and Zhu, Xiaoyong and Zheng, Bo and Cao, Liujuan},\n  journal={arXiv preprint arXiv:2605.15824},\n  year={2026}\n}\n```\n","FashionChameleon是一个面向实时和交互式人物服装视频定制的框架。它通过单GPU实现每秒23.8帧的实时生成，支持用户在视频流中即时更换人物穿着的衣物。该项目利用了定制化的I2V先验与不匹配的参考-衣物对训练单一衣物切换模型，并引入KV缓存重调度技术以实现多衣物的互动定制。此外，FashionChameleon还采用了梯度重加权DMD方法来提高长时间预测的一致性。适用于需要快速响应及高度互动性的在线试衣、虚拟换装等场景。","2026-06-11 03:57:13","CREATED_QUERY"]