[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-1917":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":23,"hasPages":23,"topics":25,"createdAt":10,"pushedAt":10,"updatedAt":28,"readmeContent":29,"aiSummary":30,"trendingCount":16,"starSnapshotCount":16,"syncStatus":15,"lastSyncTime":31,"discoverSource":32},1917,"TokenGS","nv-tlabs\u002FTokenGS","nv-tlabs","[CVPR'26] TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens","",null,"Python",209,7,4,2,0,17,25,53,51,2.71,"Apache License 2.0",false,"main",[26,27],"3d-reconstruction","novel-view-synthesis","2026-06-12 02:00:34","# TokenGS\n\n![Teaser: TokenGS results and exploration](assets\u002Ftokengs_explore.gif)\n\n**TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens** \u003Cbr>\nJiawei Ren*, Michal Tyszkiewicz*, Jiahui Huang†, Zan Gojcic† \u003Cbr>\n\\* indicates equal contribution, † indicates equal advising\n\n[**Paper**](https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.15239) · [**Project Page**](https:\u002F\u002Fresearch.nvidia.com\u002Flabs\u002Ftoronto-ai\u002Ftokengs\u002F) · [**HuggingFace**]()\n\nTokenGS predicts 3D Gaussians with a self-supervised rendering objective. An encoder–decoder stacks learnable Gaussian tokens so the number of primitives is not tied to image resolution or view count.\n\n## Installation\n\nInstall the package in editable mode (dependencies include PyTorch, gsplat, and [fused-ssim](https:\u002F\u002Fgithub.com\u002Frahul-goel\u002Ffused-ssim) via `pyproject.toml`):\n\n```bash\nuv pip install -e .\n```\n\n**Environment:** Python 3.11, CUDA 12.6+ (see `pyproject.toml` for pinned versions).\n\n**Data:** DL3DV layout, symlinks, and `dataset_kwargs` are described in **[data\u002FDATA.md](data\u002FDATA.md)**.\n\n## Evaluation\n\nPlace weights under `checkpoints\u002F` (or pass any path to `--resume`). Metrics are written to `\u003Cworkspace>\u002Fmetrics.txt`; the workspace directory is created automatically.\n\n**Example (6-view preset):**\n\n```bash\naccelerate launch --config_file acc_configs\u002Fgpu1.yaml \\\n    -m tokengs.evaluate eval_dl3dv_6view \\\n    --workspace results\u002Fdl3dv_eval\u002F6view \\\n    --resume checkpoints\u002Fdl3dv_6v.safetensors \\\n    --use_ttt_for_eval \\\n    --eval_n_media_dumps 20 \\\n```\n\nPresets `eval_dl3dv_2view` and `eval_dl3dv_4view` select the matching evaluation JSONs. Remove `--use_ttt_for_eval` to turn off test-time token tuning.\n\n**Media dumps:** `--eval_n_media_dumps N` writes PNGs, MP4s, depth vis, and PLY for the first `N` dataloader batches under `\u003Cworkspace>\u002F{images,videos,depths,gaussians}\u002F` (default `0` = metrics only).\n\n\n## Training\n\n**1. Base run** (`train_dl3dv_base` preset):\n\n```bash\naccelerate launch --config_file acc_configs\u002Fgpu8.yaml \\\n    -m tokengs.train train_dl3dv_base \\\n    --workspace workspace\u002Fdl3dv_base \\\n    --experiment_name dl3dv_base\n```\n\n**2. Finetune** from a checkpoint (presets `finetune_dl3dv_2view`, `finetune_dl3dv_4view`, `finetune_dl3dv_6view`):\n\n```bash\naccelerate launch --config_file acc_configs\u002Fgpu8.yaml \\\n    -m tokengs.train finetune_dl3dv_2view \\\n    --workspace workspace\u002Fdl3dv_2view \\\n    --experiment_name dl3dv_2view \\\n    --resume workspace\u002Fdl3dv_base\u002Fmodel.safetensors\n```\n\nSwap the subcommand for 4- or 6-view finetune presets as needed.\n\n\n## License\n\nTokenGS is released under the [Apache License 2.0](LICENSE). See [CONTRIBUTING.md](CONTRIBUTING.md) for contribution guidelines.\n\n## Citation\n\nIf you use TokenGS in your research, please cite:\n\n```bibtex\n@article{tokengs2026,\n  title={TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens},\n  author={Jiawei Ren and Michal Tyszkiewicz and Jiahui Huang and Zan Gojcic},\n  journal={Proceedings of the IEEE\u002FCVF Conference on Computer Vision and Pattern Recognition},\n  year={2026}\n}\n```\n","TokenGS是一个用于3D高斯预测的项目，通过自监督渲染目标和可学习的高斯令牌来实现从像素中解耦3D高斯预测。其核心功能包括使用编码-解码器架构堆叠可学习的高斯令牌，从而使得原语数量不受图像分辨率或视图数量的影响。技术上基于Python开发，并依赖PyTorch等库。适用于需要进行3D重建及新视角合成的研究场景，尤其是在计算机视觉领域内探索更加灵活高效的3D表示方法时。","2026-06-11 02:46:49","CREATED_QUERY"]