Generation in Generation

Research snapshot for “Generation in Generation: Fluid Co-speech Gesture Synthesis with Generative Continuous Quantization” by Jialu Li, Yifan Zhao, Xin Guo, and Jia Li.

This release follows an archival policy: the research source files are copied byte-for-byte from the experiment directory. No model, trainer, dataloader, metric, or configuration implementation has been corrected or refactored. Machine-specific paths and legacy metric behavior are therefore intentionally retained so that the saved logs remain traceable to the exact code that produced them.

Datasets and the separately licensed SMPL-X model are not redistributed.

.
├── checkpoints/             # Seven verified checkpoints, stored with Git LFS
├── configs/                 # Original experiment/evaluation YAML files
├── dataloaders/             # Original BEATv2 loader and metric implementation
├── models/                  # Original GenTau and quantization implementation
├── optimizers/              # Original optimizer/scheduler implementation
├── validation/
│   ├── RUN_PAPER_ALIGNMENT.csv
│   ├── METRIC_VALIDATION.md
│   └── raw_logs/            # Original stdout logs used by the comparison
├── train.py                 # Original training entry point
├── train_tau.py             # Original temperature-training wrapper
├── test.py                  # Original evaluation entry point
└── test_eval.py             # Original temperature-evaluation wrapper

The release checkpoint gentau_beatv2_speaker2.bin is byte-identical to the experiment checkpoint EMAGE/last_351.bin:

SHA256 014d2921efcd33c269d55834d0bc0d720cb1d580b679b98a873f628f74b88de4

The complete comparison is in validation/METRIC_VALIDATION.md, with machine-readable values in validation/RUN_PAPER_ALIGNMENT.csv.

The most important comparison is the paper's selected tau_g=0.7 setting:

These rows align to rounding precision: the largest absolute difference is 0.002208 for Speaker-2 BC and 0.000685 for cross-speaker BC.

At tau_g=1.0, the saved runs are close but not identical to the paper. The largest discrepancies are:

• Speaker-2 FGD: paper 4.943, saved run 4.979775 (+0.036775).
• Cross-speaker Diversity: paper 4.695, saved run 4.646849 (-0.048151).
• Cross-speaker BC: paper 7.296, saved run 7.205565 (-0.090435).

Thus, the tau_g=0.7 headline results are directly supported by the retained logs, while the tau_g=1.0 paper row appears to come from a different evaluation run/state than the saved log and should not be described as an exact reproduction.

Verify their hashes and Stage-1/Stage-2 codebook correspondence:

python tools/verify_release.py
(cd checkpoints && sha256sum -c SHA256SUMS)

The original environment specification is retained in requirements.txt. The original scripts contain experiment-machine paths. To rerun elsewhere, recreate the expected BEATv2, cache, SMPL-X, and checkpoint layout, or make a separate local copy in which paths are adjusted. Such local changes should not replace this archived snapshot if exact provenance is required.

Typical original entry points were:

python train_tau.py --config configs/emage_gentau_T007.yaml
python test_eval.py --config configs/emage_soft_gentau_eval.yaml --test_vq_tau 0.7

Git LFS is required before cloning or publishing:

git lfs install
git lfs pull

@article{li2026gentau,
  title   = {Generation in Generation: Fluid Co-speech Gesture Synthesis with Generative Continuous Quantization},
  author  = {Li, Jialu and Zhao, Yifan and Guo, Xin and Li, Jia},
  year    = {2026}
}

Code is distributed under Apache-2.0. Dataset, pretrained third-party model, and SMPL-X licenses remain subject to their respective terms.