[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-72513":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":27,"readmeContent":28,"aiSummary":29,"trendingCount":16,"starSnapshotCount":16,"syncStatus":30,"lastSyncTime":31,"discoverSource":32},72513,"JiT","LTH14\u002FJiT","LTH14","PyTorch implementation of JiT https:\u002F\u002Farxiv.org\u002Fabs\u002F2511.13720","",null,"Python",2387,162,13,51,0,8,26,85,24,96.14,"MIT License",false,"main",true,[],"2026-06-12 04:01:06","## Just image Transformer (JiT) for Pixel-space Diffusion\n\n[![arXiv](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv%20paper-2511.13720-b31b1b.svg)](https:\u002F\u002Farxiv.org\u002Fabs\u002F2511.13720) \n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"demo\u002Fvisual.jpg\" width=\"100%\">\n\u003C\u002Fp>\n\n\nThis is a PyTorch\u002FGPU re-implementation of the paper [Back to Basics: Let Denoising Generative Models Denoise](https:\u002F\u002Farxiv.org\u002Fabs\u002F2511.13720):\n\n```\n@article{li2025jit,\n  title={Back to Basics: Let Denoising Generative Models Denoise},\n  author={Li, Tianhong and He, Kaiming},\n  journal={arXiv preprint arXiv:2511.13720},\n  year={2025}\n}\n```\n\nJiT adopts a minimalist and self-contained design for pixel-level high-resolution image diffusion. \nThe original implementation was in JAX+TPU. This re-implementation is in PyTorch+GPU.\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"demo\u002Fjit.jpg\" width=\"40%\">\n\u003C\u002Fp>\n\n### Dataset\nDownload [ImageNet](http:\u002F\u002Fimage-net.org\u002Fdownload) dataset, and place it in your `IMAGENET_PATH`.\n\n### Installation\n\nDownload the code:\n```\ngit clone https:\u002F\u002Fgithub.com\u002FLTH14\u002FJiT.git\ncd JiT\n```\n\nA suitable [conda](https:\u002F\u002Fconda.io\u002F) environment named `jit` can be created and activated with:\n\n```\nconda env create -f environment.yaml\nconda activate jit\n```\n\nIf you get ```undefined symbol: iJIT_NotifyEvent``` when importing ```torch```, simply\n```\npip uninstall torch\npip install torch==2.5.1 --index-url https:\u002F\u002Fdownload.pytorch.org\u002Fwhl\u002Fcu124\n```\nCheck this [issue](https:\u002F\u002Fgithub.com\u002Fconda\u002Fconda\u002Fissues\u002F13812#issuecomment-2071445372) for more details.\n\n### Training\nThe below training scripts have been tested on 8 H200 GPUs.\n\nExample script for training JiT-B\u002F16 on ImageNet 256x256 for 600 epochs:\n```\ntorchrun --nproc_per_node=8 --nnodes=1 --node_rank=0 \\\nmain_jit.py \\\n--model JiT-B\u002F16 \\\n--proj_dropout 0.0 \\\n--P_mean -0.8 --P_std 0.8 \\\n--img_size 256 --noise_scale 1.0 \\\n--batch_size 128 --blr 5e-5 \\\n--epochs 600 --warmup_epochs 5 \\\n--gen_bsz 128 --num_images 50000 --cfg 2.9 --interval_min 0.1 --interval_max 1.0 \\\n--output_dir ${OUTPUT_DIR} --resume ${OUTPUT_DIR} \\\n--data_path ${IMAGENET_PATH} --online_eval\n```\n\nExample script for training JiT-B\u002F32 on ImageNet 512x512 for 600 epochs:\n```\ntorchrun --nproc_per_node=8 --nnodes=1 --node_rank=0 \\\nmain_jit.py \\\n--model JiT-B\u002F32 \\\n--proj_dropout 0.0 \\\n--P_mean -0.8 --P_std 0.8 \\\n--img_size 512 --noise_scale 2.0 \\\n--batch_size 128 --blr 5e-5 \\\n--epochs 600 --warmup_epochs 5 \\\n--gen_bsz 128 --num_images 50000 --cfg 2.9 --interval_min 0.1 --interval_max 1.0 \\\n--output_dir ${OUTPUT_DIR} --resume ${OUTPUT_DIR} \\\n--data_path ${IMAGENET_PATH} --online_eval\n```\n\nExample script for training JiT-H\u002F16 on ImageNet 256x256 for 600 epochs:\n```\ntorchrun --nproc_per_node=8 --nnodes=1 --node_rank=0 \\\nmain_jit.py \\\n--model JiT-H\u002F16 \\\n--proj_dropout 0.2 \\\n--P_mean -0.8 --P_std 0.8 \\\n--img_size 256 --noise_scale 1.0 \\\n--batch_size 128 --blr 5e-5 \\\n--epochs 600 --warmup_epochs 5 \\\n--gen_bsz 128 --num_images 50000 --cfg 2.2 --interval_min 0.1 --interval_max 1.0 \\\n--output_dir ${OUTPUT_DIR} --resume ${OUTPUT_DIR} \\\n--data_path ${IMAGENET_PATH} --online_eval\n```\n\n### Evaluation\n\nPyTorch pre-trained models are available [here](https:\u002F\u002Fwww.dropbox.com\u002Fscl\u002Ffo\u002F3ken1avtsd81ip67b9qpi\u002FAK218ZNvXKSv74igVvht4PQ?rlkey=14gjrblmljewpl6ygxzlr3njm&st=ffkl77al&dl=0).\n\nEvaluate pre-trained JiT-B:\n```\ntorchrun --nproc_per_node=8 --nnodes=1 --node_rank=0 \\\nmain_jit.py \\\n--model JiT-B\u002F16 (or JiT-B\u002F32) \\\n--img_size 256 (or 512) --noise_scale 1.0 (or 2.0) \\\n--gen_bsz 256 --num_images 50000 --cfg 3.0 --interval_min 0.1 --interval_max 1.0 \\\n--output_dir ${CKPT_DIR} --resume ${CKPT_DIR} \\\n--data_path ${IMAGENET_PATH} --evaluate_gen\n```\n\nEvaluate pre-trained JiT-L:\n```\ntorchrun --nproc_per_node=8 --nnodes=1 --node_rank=0 \\\nmain_jit.py \\\n--model JiT-L\u002F16 (or JiT-L\u002F32) \\\n--img_size 256 (or 512) --noise_scale 1.0 (or 2.0) \\\n--gen_bsz 256 --num_images 50000 --cfg 2.4 (or 2.5) --interval_min 0.1 --interval_max 1.0 \\\n--output_dir ${CKPT_DIR} --resume ${CKPT_DIR} \\\n--data_path ${IMAGENET_PATH} --evaluate_gen\n```\n\nEvaluate pre-trained JiT-H:\n```\ntorchrun --nproc_per_node=8 --nnodes=1 --node_rank=0 \\\nmain_jit.py \\\n--model JiT-H\u002F16 (or JiT-H\u002F32) \\\n--img_size 256 (or 512) --noise_scale 1.0 (or 2.0) \\\n--gen_bsz 256 --num_images 50000 --cfg 2.2 (or 2.3) --interval_min 0.1 --interval_max 1.0 \\\n--output_dir ${CKPT_DIR} --resume ${CKPT_DIR} \\\n--data_path ${IMAGENET_PATH} --evaluate_gen\n```\n\nWe use a customized [```torch-fidelity```](https:\u002F\u002Fgithub.com\u002FLTH14\u002Ftorch-fidelity)\nto evaluate FID and IS against a reference image folder or statistics. You can use ```prepare_ref.py```\nto prepare the reference image folder, or directly use our pre-computed reference stats\nunder ```fid_stats```.\n\n### Acknowledgements\n\nWe thank Google TPU Research Cloud (TRC) for granting us access to TPUs, and the MIT\nORCD Seed Fund Grants for supporting GPU resources.\n\n### Contact\n\nIf you have any questions, feel free to contact me through email (tianhong@mit.edu). Enjoy!\n","JiT是一个基于PyTorch实现的像素级高分辨率图像扩散模型，原始论文为《Back to Basics: Let Denoising Generative Models Denoise》。该项目采用简洁且自包含的设计，特别适用于在GPU环境下进行高效的图像生成与处理任务。其核心功能包括支持多种分辨率（如256x256、512x512）下的图像生成，并提供了详细的训练脚本示例，方便用户根据自身需求调整参数配置。JiT适合于需要高质量图像生成的应用场景，例如艺术创作、数据增强或研究实验等。此外，项目遵循MIT许可证，鼓励开源社区的参与和贡献。",2,"2026-06-11 03:42:21","high_star"]