---
license: mit
task_categories:
  - other
tags:
  - motion-generation
  - diffusion-models
  - cross-skeleton
  - retargeting
  - anonymous-submission
size_categories:
  - 1K<n<10K
---

# anytop-checkpoints

Pre-trained checkpoints accompanying the anonymous NeurIPS 2026 submission
**"Why Cross-Skeleton Retargeting Is Non-Identifiable: Structural Limits of
Generative Motion Models"**.

This repository contains a single tarball, `anytop-checkpoints.tar.gz`
(6.5 GB compressed, 7.2 GB extracted), bundling the trained model weights
for the 14-method comparison reported in the paper.

## Contents

After extraction (`tar -xzvf anytop-checkpoints.tar.gz`):

```
anytop-checkpoints/
  README.md                       (per-method usage notes)
  anytop_v5/                      AnyTop transductive (model000175000.pt, 34 MB)
  ace_primary_70/                 ACE-T transductive (ckpt_final.pt, 83 MB)
  ace_inductive_60/               ACE-I inductive (ckpt_final.pt, 83 MB)
  moreflow_primary_70/            MoReFlow-T (ckpt_final.pt, 79 MB)
  moreflow_inductive_60/          MoReFlow-I (ckpt_final.pt, 79 MB)
  anchor_label_flow/              AL-Flow (ckpt_final.pt, 86 MB)
  anchor_label_flow_src/          AL-Flow-Src (ckpt_final.pt, 88 MB)
  anchor_label_flow_src_g/        AL-Flow-Src-G (ckpt_final.pt, 87 MB)
  dpg_sb_v3/                      DPG-SB-v3 (final.pt 62 MB + z_stats.pt)
  moreflow_vqvae/                 71 per-skeleton VQ-VAE encoders (~91 MB each)
                                  Required for ACE / MoReFlow / DPG-SB-v3 inference
  moreflow_caches/                Pre-extracted MoReFlow latents
                                  (only required to retrain DPG-SB-v3 from scratch)
```

## Verification

```
SHA-256: fc15f536510014a7bb953aa62cd24358de6af11f0d92f0f8360e2ec44b8251e2
```

```bash
sha256sum -c <<< "fc15f536510014a7bb953aa62cd24358de6af11f0d92f0f8360e2ec44b8251e2  anytop-checkpoints.tar.gz"
```

## How to use

The companion code repository contains training scripts, evaluation runners,
and the SIF metric implementation. After downloading and extracting this
tarball into the code repo's `save/` directory, every per-method invocation
in the code repo's `REPRODUCE.md` § 4 will work without retraining.

```bash
# From the code repo root:
mkdir -p save
tar -xzvf anytop-checkpoints.tar.gz
mv anytop-checkpoints/* save/
rmdir anytop-checkpoints

# Verify with one method:
python -m eval.baselines.run_anytop_v5 \
    --ckpt save/anytop_v5/model000175000.pt \
    --manifest eval/benchmark_v3/queries_sif/manifest.json \
    --output_dir results/anytop_sif
```

## License

MIT, matching the upstream AnyTop release.

## Anonymity

Every `args.json` file in the package has been swept for personal paths,
author names, and W&B entity strings. Checkpoint binaries (`*.pt`) contain
only PyTorch model and optimiser state.