---
license: mit
tags:
- text-to-motion
- motion-generation
- humanml3d
- codeflow
- rvq
- pytorch
---

# CodeFlow HumanML3D

Public HumanML3D checkpoint bundle for Part-Structured CodeFlow (PS-CF).

## Contents

- `codeflow/codeflow_hml3d_best_top3_ema.pt`: inference-only EMA checkpoint exported from the HumanML3D best-Top3 CodeFlow run.
- `rvq/part_vq_hml3d_overlap_best_top3.pth`: frozen part-aware RVQ tokenizer checkpoint used by CodeFlow.
- `rvq/skeleton_partition.json`: six-part overlap partition for the RVQ tokenizer.
- `stats/mean.npy`, `stats/std.npy`: HumanML3D normalization statistics used by the RVQ tokenizer.
- `metadata/codeflow_options.json`: sanitized training/options metadata.
- `metadata/best_metrics.json`: training-time full-eval best checkpoint record.
- `metadata/rvq_metrics.md`: RVQ training-time metrics summary.

## Released CodeFlow Checkpoint

Selector: training-time full-eval best Top3 on HumanML3D test.

- epoch: 290
- step: 111070
- Top3: 0.873060344828
- FID: 0.058189920627
- architecture: part-structured CodeFlow, `part_hidden_dim=192`, `hidden_size=1152`, `depth_double=6`, `depth_single=12`, `dropout=0.05`
- evaluation setting: test split, 96 ODE steps, CFG=6.0, EMA weights, seed 42

## Notes

This release does not include HumanML3D data, CLIP weights, GloVe, evaluator checkpoints, KIT checkpoints, or optimizer/scaler training states.
Use the GitHub repository instructions for inference and evaluation commands.