--- license: other license_name: babel-amass-non-commercial license_link: https://babel.is.tue.mpg.de/license.html pretty_name: Triangular Resampling data (BABEL 263D) and checkpoints tags: - motion-generation - text-to-motion - human-motion - babel - humanml3d - diffusion --- # Triangular Resampling: data and checkpoints Data and checkpoints for **Triangular Resampling for Long-Horizon Motion Generation** ([arXiv:2609.34697](https://arxiv.org/abs/2609.34697)). Code: [github.com/KunhangL/TriangularResampling](https://github.com/KunhangL/TriangularResampling). This repository holds the BABEL data and the FloodDiffusion 2 (FD2) training metadata used in the paper. These are the exact files of the paper's runs. HumanML3D is not included; it comes with the [FloodDiffusion release](https://huggingface.co/ShandaAI/FloodDiffusionDownloads), whose `HumanML3D.zip` is identical to the HumanML3D copy used in the paper. The repository also holds FD2 post-training checkpoints, described under [`checkpoints/`](#checkpoints). | File | Extract into | Contents | SHA-256 | | --- | --- | --- | --- | | `BABEL.zip` | `raw_data/` | 263D BABEL features, annotations, FD1 VAE tokens, splits | `681fdd86ebc70d24bec18b1b83e04f0070a8ae8fc3cd3a33ee7f1cd7ba2dfdcf` | | `manifests.zip` | `raw_data/` | FD2 training membership | `71b2b95ff2c072500562aa8ebe2b42d03dbbd4102b2d07f7e9c83e566d88cc32` | ## Download From a clone of the code repository (checks the SHA-256 of both archives): ```bash python -m common.download_data # both archives into raw_data/ python -m methods.fd1.download_assets --with-dataset # FD1: deps, checkpoints, HumanML3D and BABEL python -m methods.fd2.download_assets --with-deps --with-data # FD2: checkpoint, deps, BABEL, manifests ``` Or directly: ```python import zipfile from huggingface_hub import hf_hub_download for name in ("BABEL.zip", "manifests.zip"): path = hf_hub_download("KunhangLI/TriangularResampling", name) zipfile.ZipFile(path).extractall("raw_data") ``` The checkpoints, into `checkpoints/` of the code repository: ```bash hf download KunhangLI/TriangularResampling \ checkpoints/fd2_tr_step200.ckpt \ checkpoints/fd2_tr_dmd_step400.ckpt \ --local-dir . ``` ## `BABEL.zip` ``` BABEL/ ├── new_joint_vecs/.npy 5,341 sequences, float32 (T, 263), 20 fps ├── texts/.txt 4,877 annotation files ├── TOKENS_20251030_085836_vae_wan_z4/.npy 5,341 VAE token files, float32 (ceil(T/4), 4) ├── train.txt 3,985 ids (BABEL train split) ├── val.txt 1,356 ids (BABEL validation split) └── test_min.txt 10 ids (validation sequences used for rendering) ``` - **Features.** Every BABEL sequence (train and validation) in the 263-dimensional HumanML3D motion representation at 20 fps, one file per BABEL sequence id, 25 to 5,272 frames. Sequences are complete, not cut into segments. - **Annotations.** HumanML3D-style text files with one line per labelled BABEL segment: `caption#word/POS ...#start_sec#end_sec`. 464 sequences have no annotation; the code skips them. - **Tokens.** The latents of the official FloodDiffusion VAE (`vae_1d_z4_step=300000.ckpt`), 4 frames per token, as read by the FD1 configs. They are the tokens of the paper's FD1 runs, so FD1 can be trained without re-encoding. - **Splits.** `train.txt` and `val.txt` follow BABEL's train / validation split. Training uses the 3,640 annotated training sequences. The validation sequences serve as reference motions for the paper's time-varying-text evaluation, whose frozen prompt manifest is `common/evaluation/manifests/babel_text_streams_256.jsonl` in the code repository. ## `manifests.zip` ``` manifests/ └── humanml_babel.json FD2 training membership ``` - **`humanml_babel.json`** lists every HumanML3D and BABEL sample of the FD2 run, with paths relative to the code repository root. It has 25,966 training samples, 4,384 validation samples (the full HumanML3D test split) and 18 rendering samples. Its SHA-256, `b27b204b48915114cd69ba016f25640e75ba809f94c496c27184cb6f63750bac`, is pinned in `methods/fd2/configs/humanml_babel_tr.yaml`. Running `methods/fd2/prepare_data.py` from the repository root rebuilds the same file. Its `problems` field lists two HumanML3D test features with non-finite values (`007975`, `M007975`), which FD2 validation skips. ## `checkpoints/` FD2 post-training checkpoints for HumanML3D + BABEL (263D features at 20 fps). Both start from the official FD2 `humanml3d_babel_fk_200k` checkpoint ([AlayaLab/FloodDiffusion2](https://huggingface.co/AlayaLab/FloodDiffusion2)). | File | Model | Config in the code repository | SHA-256 | | --- | --- | --- | --- | | `checkpoints/fd2_tr_step200.ckpt` | FD2 + TR, 200 post-training steps | `methods/fd2/configs/humanml_babel_tr.yaml` | `4aa7fc36cea05b6e1ac72e78bcd225a75ae6835a172e294537414545e546f104` | | `checkpoints/fd2_tr_dmd_step400.ckpt` | FD2 + TR + DMD, 400 post-training steps | `methods/fd2/configs/humanml_babel_tr_dmd.yaml` | `bd505473cd8377d6c86e27c98fc573d737bfed969254bfa1708ad185a7214d16` | - **Format.** Inference exports of about 686 MB each, with three entries: `state_dict` (bfloat16), `ema_state` (float32 EMA weights, the ones generation uses) and `global_step`. They hold no optimizer state, so training cannot be resumed from them. - **Use.** Pass the file and its config to `methods/fd2/training/generate.py`: ```bash python -m methods.fd2.training.generate --config methods/fd2/configs/humanml_babel_tr.yaml \ --checkpoint checkpoints/fd2_tr_step200.ckpt \ --manifest common/evaluation/manifests/humanml3d_prompts_256_uniform_seed1234.jsonl \ --output outputs/eval/fd2_tr_fixed/generated --method fd2_tr --seed 42 ``` The code repository's README lists the 120-second evaluation scores of both checkpoints. ## License These files are derived from [BABEL](https://babel.is.tue.mpg.de/) (Punnakkal et al., CVPR 2021), which builds on [AMASS](https://amass.is.tue.mpg.de/) (Mahmood et al., ICCV 2019). Only extracted motion features, text annotations, model latents and id lists are provided. No raw AMASS data (SMPL parameters or meshes) is distributed. The data is for **non-commercial scientific research only**. By using it you agree to the [BABEL license](https://babel.is.tue.mpg.de/license.html) and the [AMASS license](https://amass.is.tue.mpg.de/license.html). If you need the raw motion data or a commercial license, register with and contact the dataset owners. The checkpoints are trained on HumanML3D and BABEL, starting from the official FD2 checkpoint. They are released under the same terms: non-commercial scientific research only. ## Citation ```bibtex @article{li2026triangular, title = {Triangular Resampling for Long-Horizon Motion Generation}, author = {Li, Kunhang and Cai, Yiyi and Zhang, Xiangyue and Tu, Fangyuan and Wu, Yuhan and Wang, Zhixiang and Zhang, Kaipeng and Liu, Haiyang}, journal = {arXiv preprint arXiv:2609.34697}, year = {2026} } @inproceedings{punnakkal2021babel, title = {{BABEL}: Bodies, Action and Behavior with English Labels}, author = {Punnakkal, Abhinanda R. and Chandrasekaran, Arjun and Athanasiou, Nikos and Quiros-Ramirez, Alejandra and Black, Michael J.}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, year = {2021} } @inproceedings{mahmood2019amass, title = {{AMASS}: Archive of Motion Capture as Surface Shapes}, author = {Mahmood, Naureen and Ghorbani, Nima and Troje, Nikolaus F. and Pons-Moll, Gerard and Black, Michael J.}, booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)}, year = {2019} } ```