{"slug": "test-time-training-3d-reconstruction", "title": "Test-time training 3D reconstruction", "summary": "Researchers released TTT3R, an open-source 3D reconstruction method that treats reconstruction as test-time training. The project provides code, pretrained checkpoints, and inference demos for converting video or image sequences into 3D models.", "body_md": "## ttt3r.mp4\n\n- Clone TTT3R.\n\n```\ngit clone https://github.com/Inception3D/TTT3R.git\ncd TTT3R\n```\n\n- Create the environment.\n\n```\nconda create -n ttt3r python=3.11 cmake=3.14.0\nconda activate ttt3r\nconda install pytorch torchvision pytorch-cuda=12.1 -c pytorch -c nvidia  # use the correct version of cuda for your system\npip install -r requirements.txt\n# issues with pytorch dataloader, see https://github.com/pytorch/pytorch/issues/99625\nconda install 'llvm-openmp<16'\n# for evaluation\npip install evo\npip install open3d\n```\n\n- Compile the cuda kernels for RoPE (as in CroCo v2).\n\n```\ncd src/croco/models/curope/\npython setup.py build_ext --inplace\ncd ../../../../\n```\n\nCUT3R provide checkpoints trained on 4-64 views: [ cut3r_512_dpt_4_64.pth](https://drive.google.com/file/d/1Asz-ZB3FfpzZYwunhQvNPZEUA8XUNAYD/view?usp=drive_link).\n\nTo download the weights, run the following commands:\n\n```\ncd src\ngdown --fuzzy https://drive.google.com/file/d/1Asz-ZB3FfpzZYwunhQvNPZEUA8XUNAYD/view?usp=drive_link\ncd ..\n```\n\nTo run the inference demo, you can use the following command:\n\n```\n# input can be a folder or a video\n# the following script will run inference with TTT3R and visualize the output with viser on port 8080\nCUDA_VISIBLE_DEVICES=6 python demo.py --model_path MODEL_PATH --size 512 \\\n    --seq_path SEQ_PATH --output_dir OUT_DIR --port 8080 \\\n    --model_update_type ttt3r --frame_interval 1 --reset_interval 100 \\\n    --downsample_factor 1000 --vis_threshold 5.0\n\n# Example:\nCUDA_VISIBLE_DEVICES=6 python demo.py --model_path src/cut3r_512_dpt_4_64.pth --size 512 \\\n    --seq_path examples/westlake.mp4 --output_dir tmp/taylor --port 8080 \\\n    --model_update_type ttt3r --frame_interval 1 --reset_interval 100 \\\n    --downsample_factor 100 --vis_threshold 6.0\n\nCUDA_VISIBLE_DEVICES=6 python demo.py --model_path src/cut3r_512_dpt_4_64.pth --size 512 \\\n    --seq_path examples/taylor.mp4 --output_dir tmp/taylor --port 8080 \\\n    --model_update_type ttt3r --frame_interval 1 --reset_interval 50 \\\n    --downsample_factor 100 --vis_threshold 10.0\n```\n\nOutput results will be saved to `output_dir`\n\n.\n\nPlease refer to the [eval.md](/Inception3D/TTT3R/blob/main/eval/eval.md) for more details.\n\nOur code is based on the following awesome repositories:\n\nWe thank the authors for releasing their code!\n\nIf you find our work useful, please cite:\n\n```\n@article{chen2025ttt3r,\n    title={TTT3R: 3D Reconstruction as Test-Time Training},\n    author={Chen, Xingyu and Chen, Yue and Xiu, Yuliang and Geiger, Andreas and Chen, Anpei},\n    journal={arXiv preprint arXiv:2509.26645},\n    year={2025}\n    }\n```\n\n", "url": "https://wpnews.pro/news/test-time-training-3d-reconstruction", "canonical_source": "https://github.com/Inception3D/TTT3R", "published_at": "2026-07-16 00:11:55+00:00", "updated_at": "2026-07-16 00:25:22.757080+00:00", "lang": "en", "topics": ["computer-vision", "machine-learning", "artificial-intelligence"], "entities": ["Inception3D", "TTT3R", "CUT3R", "CroCo v2", "Xingyu Chen", "Yue Chen", "Yuliang Xiu", "Andreas Geiger"], "alternates": {"html": "https://wpnews.pro/news/test-time-training-3d-reconstruction", "markdown": "https://wpnews.pro/news/test-time-training-3d-reconstruction.md", "text": "https://wpnews.pro/news/test-time-training-3d-reconstruction.txt", "jsonld": "https://wpnews.pro/news/test-time-training-3d-reconstruction.jsonld"}}