Views
No views yet
pyproject.toml):uv pip install -e .pyproject.toml for pinned versions).dataset_kwargs are described in data/DATA.md.checkpoints/ (or pass any path to --resume). Metrics are written to <workspace>/metrics.txt; the workspace directory is created automatically.| Checkpoint | Eval preset |
|---|---|
dl3dv_2v.safetensors | eval_dl3dv_2view |
dl3dv_4v.safetensors | eval_dl3dv_4view |
dl3dv_6v.safetensors | eval_dl3dv_6view |
1accelerate launch --config_file acc_configs/gpu1.yaml \
2 -m tokengs.evaluate eval_dl3dv_6view \
3 --workspace results/dl3dv_eval/6view \
4 --resume checkpoints/dl3dv_6v.safetensors \
5 --use_ttt_for_eval \
6 --eval_n_media_dumps 20 \eval_dl3dv_2view and eval_dl3dv_4view select the matching evaluation JSONs. Remove --use_ttt_for_eval to turn off test-time token tuning.--eval_n_media_dumps N writes PNGs, MP4s, depth vis, and PLY for the first N dataloader batches under <workspace>/{images,videos,depths,gaussians}/ (default 0 = metrics only).train_dl3dv_base preset):1accelerate launch --config_file acc_configs/gpu8.yaml \
2 -m tokengs.train train_dl3dv_base \
3 --workspace workspace/dl3dv_base \
4 --experiment_name dl3dv_basefinetune_dl3dv_2view, finetune_dl3dv_4view, finetune_dl3dv_6view):1accelerate launch --config_file acc_configs/gpu8.yaml \
2 -m tokengs.train finetune_dl3dv_2view \
3 --workspace workspace/dl3dv_2view \
4 --experiment_name dl3dv_2view \
5 --resume workspace/dl3dv_base/model.safetensorscheckpoints_26.06/kubric_dyn.safetensors. Point data/kubric at the Kubric multi-view dump:
1data/kubric/
2 v0/
3 <scene>/
4 output_000.tar
5 output_001.tar
6 ...
7 v1/
8 v2/
9 ...v0, v1, v2, ... folders are data splits. Each output_{view:03d}.tar contains metadata.json, rgba_{frame:05d}.png, and depth_{frame:05d}.tiff. The dynamic presets use pointmap camera scaling, so input-frame depth TIFFs are loaded for scale normalization.0.3 for 5K steps, then linearly decays to 0 over the next 10K steps.1accelerate launch --config_file acc_configs/gpu8.yaml \
2 -m tokengs.train finetune_dl3dv_kubric_dyn_release \
3 --workspace workspace/kubric_dyn \
4 --experiment_name kubric_dyn \
5 --resume checkpoints/dl3dv_base.safetensors--workspace only when rendering a finetuned workspace with its own config.yaml.1python scripts/render_kubric_dyn.py \
2 --preset finetune_dl3dv_kubric_dyn_release \
3 --ckpt checkpoints_26.06/kubric_dyn.safetensors \
4 --kubric_root data/kubric \
5 --out_dir results/kubric_dyn_renders \
6 --scene_idx 0 \
7 --n_scenes 4 \
8 --fps 81@article{tokengs2026,
2 title={TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens},
3 author={Jiawei Ren and Michal Tyszkiewicz and Jiahui Huang and Zan Gojcic},
4 journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
5 year={2026}
6}