[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-1049":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":8,"htmlUrl":8,"language":9,"languages":8,"totalLinesOfCode":8,"stars":10,"forks":11,"watchers":12,"openIssues":13,"contributorsCount":14,"subscribersCount":14,"size":14,"stars1d":15,"stars7d":16,"stars30d":17,"stars90d":14,"forks30d":14,"starsTrendScore":18,"compositeScore":19,"rankGlobal":8,"rankLanguage":8,"license":20,"archived":21,"fork":21,"defaultBranch":22,"hasWiki":23,"hasPages":21,"topics":24,"createdAt":8,"pushedAt":8,"updatedAt":25,"readmeContent":26,"aiSummary":27,"trendingCount":14,"starSnapshotCount":14,"syncStatus":12,"lastSyncTime":28,"discoverSource":29},1049,"FD-Loss","Jiawei-Yang\u002FFD-Loss","Jiawei-Yang",null,"Python",525,12,2,3,0,7,9,50,21,7.34,"MIT License",false,"main",true,[],"2026-06-12 02:00:22","## Representation Fréchet Loss for Visual Generation\n\n[![arXiv](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2604.28190-b31b1b.svg)](https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.28190)\n[![Hugging Face checkpoints](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FHuggingFace-checkpoints-ffcc00.svg)](https:\u002F\u002Fhuggingface.co\u002Fjjiaweiyang\u002FFD-Loss)\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"docs\u002Ffd_loss_dynamic.svg\" width=\"760\" alt=\"FD-Loss training dynamics\">\n\u003C\u002Fp>\n\nThis is a PyTorch\u002FGPU implementation of the paper:\n[Representation Fréchet Loss for Visual Generation](https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.28190).\n\n```bibtex\n@article{yang2026fdloss,\n  title={Representation Fréchet Loss for Visual Generation},\n  author={Yang, Jiawei and Geng, Zhengyang and Ju, Xuan and Tian, Yonglong and Wang, Yue},\n  journal={arXiv:2604.28190},\n  url={https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.28190},\n  year={2026}\n}\n```\n\nFD-Loss post-trains visual generators by matching generated-image feature\ndistributions to real-image feature distributions in frozen representation spaces.\nThis repository includes training, released-checkpoint evaluation, reference\nstatistics utilities, and scripts for the ImageNet experiments.\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"docs\u002Fvisual.png\" width=\"760\" alt=\"FD-Loss visual overview\">\n\u003C\u002Fp>\n\n### Dataset\n\nDownload ImageNet and place it in your `DATA_ROOT` using the standard\n`ImageFolder` layout:\n\n```bash\nexport DATA_ROOT=\u002Fpath\u002Fto\u002Fimagenet\n```\n\n### Installation\n\nDownload the code:\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FJiawei-Yang\u002FFD-Loss.git\ncd FD-Loss\n```\n\nCreate and activate a conda environment:\n\n```bash\nconda create -n fdloss python=3.11 -y\nconda activate fdloss\n\npip install --upgrade pip\npip install torch==2.6.0 torchvision==0.21.0 --index-url https:\u002F\u002Fdownload.pytorch.org\u002Fwhl\u002Fcu124\npip install -r requirements.txt\npip install -U huggingface_hub\n```\n\n### Checkpoints And Statistics\n\nReleased checkpoints and data files are hosted on\n[Hugging Face](https:\u002F\u002Fhuggingface.co\u002Fjjiaweiyang\u002FFD-Loss).\n\n```bash\nhf download jjiaweiyang\u002FFD-Loss \\\n  --local-dir . \\\n  --include \"checkpoints\u002F**\u002F*.pth\" \\\n  --include \"data\u002F**\"\n\npython scripts\u002Fextract_paper_ref_stats.py\n```\n\nSee [scripts\u002FREADME.md](scripts\u002FREADME.md) for the asset layout and lighter\ndownload options.\n\n### Evaluation\n\nEvaluate the released FD-SIM models:\n\n```bash\nPRESET=pMF_H_256 \\\nCKPT_PATH=checkpoints\u002Fpost-trained\u002FpMF-H_FD-SIM.pth \\\nGPUS_PER_NODE=8 \\\nbash scripts\u002Fevaluate_released_ckpt.sh\n\nPRESET=JiT_H \\\nCKPT_PATH=checkpoints\u002Fpost-trained\u002FJiT-H_FD-SIM.pth \\\nGPUS_PER_NODE=8 \\\nbash scripts\u002Fevaluate_released_ckpt.sh\n\nPRESET=iMF_XL \\\nCKPT_PATH=checkpoints\u002Fpost-trained\u002FiMF-XL_FD-SIM.pth \\\nGPUS_PER_NODE=8 \\\nbash scripts\u002Fevaluate_released_ckpt.sh\n```\n\nAdditional presets and smoke-test settings are documented in\n[scripts\u002FREADME.md](scripts\u002FREADME.md).\n\n### Training\n\nTraining starts from the released base checkpoints:\n\n```bash\nexport CKPT_ROOT=.\u002Fcheckpoints\u002Fbase\n```\n\nThe experiment scripts under [scripts\u002F](scripts\u002F) reproduce the Table 1\nablations, Table 2 JiT repurposing, and Table 3 scalability runs. For example:\n\n```bash\nbash scripts\u002Ftable_1a_queue_size.sh\nbash scripts\u002Ftable_2_repurpose_jit_L.sh\nMODEL_SIZE=L RES=256 bash scripts\u002Ftable_3_pMF.sh\nMODEL_SIZE=XL bash scripts\u002Ftable_3_iMF.sh\nMODEL_SIZE=H bash scripts\u002Ftable_3_JiT.sh\n```\n\n### License\n\nThis project is released under the MIT license. See [LICENSE](LICENSE) for details.\n\nIf you have any questions, feel free to contact me through email\n([yangjiaw@usc.edu](mailto:yangjiaw@usc.edu)).\n","FD-Loss 是一个基于 PyTorch 和 GPU 的实现，用于通过匹配生成图像和真实图像在冻结表示空间中的特征分布来后训练视觉生成器。该项目的核心功能是提供了一种称为 Representation Fréchet Loss 的损失函数，该方法能够有效提升生成模型的性能。技术上，它支持 ImageNet 数据集上的实验，并提供了预训练模型、评估脚本以及参考统计工具等资源。此项目适用于需要改进或研究高质量图像生成算法的研究者及开发者，特别是在计算机视觉和深度学习领域内对生成对抗网络（GANs）或其他类型生成模型感兴趣的用户。","2026-06-11 02:41:18","CREATED_QUERY"]