PAM — Public release¶

Development line: project:pam · thread public-release
Last event: 2026-03-26 · 1 dated since 2026-03-26 · Researched: 2026-09-06 · confidence: medium

What it is¶

PAM is a research pipeline for teams creating hand-object interaction data from simulation.

• Pose builds a trajectory from start and goal MANO poses, 6-DoF object pose, and object mesh.
• Appearance generates the first frame from depth, segmentation, and hand keypoints.
• Motion turns conditions and the chosen first frame into a video clip.

Authors report 301,1 s for 40 frames and a 41,4 GB peak on NVIDIA H20. This is an offline generator of synthetic HOI data, not a prompt-only video endpoint.

Development line¶

• 2026-03-26 — PAM public project resources were made available. The project website, code, and weight collection were linked together. The primary source confirms no new code or weight release on this day.

What changed¶

• 2025-12-04 — PAM: Hugging Face added a four-part all-condition motion checkpoint for OakInk2.
• 2026-03-23 — PAM: arXiv v1 presented the three-stage pose–appearance–motion approach.
• 2026-03-25 — PAM: arXiv v2 published results on DexYCB and OAKINK2 before the primary mention date.
• 2026-03-26 — PAM: The project website, code, and weight collection were linked together, but the primary source confirms no new code or weight release on this day.
• 2026-04-02 — PAM: arXiv v3 updated the entry; the entry comment notes acceptance to CVPR 2026.

How to use this¶

1. Check task fit first: PAM takes simulated HOI geometry and poses, not a single text prompt. — https://arxiv.org/html/2603.22193v2
2. Pick and pin a dataset-specific artifact, such as dexycb_all_cond or oakink2_all_cond, instead of treating the Hub root as a single checkpoint. — https://huggingface.co/GasaiYU/PAM/tree/main
3. Deploy the environment from environment.yaml, use CUDA 12.1, and install manopth as required by the official code. — https://github.com/GasaiYU/PAM
4. Prepare DexYCB or OakInk2: video, foreground-masked depth, segmentation, hand keypoints, captions, and filelists. — https://github.com/GasaiYU/PAM
5. For sim-to-real, pass rendered GraspXL RGB/normal/depth/segmentation/keypoints, generate and pick the first frame, then run the motion stage via testing/evaluation_sim.py. — https://github.com/GasaiYU/PAM

Best practices¶

• Do not take the auto-generated prompt-only Diffusers example on the Hub as a verified recipe: use separate artifacts and input conditions from the authors' pipeline. — https://huggingface.co/GasaiYU/PAM/tree/main
• Keep all three conditions: depth, segmentation, and hand keypoints. The authors' ablation shows the best result when combining them. — https://arxiv.org/html/2603.22193v2
• Check geometry and filter the first frame before the motion stage: contact errors and a poor reference frame turn into physically implausible video and temporal flicker. — https://arxiv.org/html/2603.22193v3
• Plan the run as a heavy offline workload: one full 40-frame pass took 301,1 s on H20 in the paper, and the appearance stage reached 41,4 GB. — https://arxiv.org/html/2603.22193v3

Superseded by this¶

• 2026-04-02 — arXiv v1 (2026-03-23) and v2 (2026-03-25) are replaced by the current v3 for citation and checking paper status.

Still unknown¶

• The primary source provides no verifiable timestamp for the state of the project website, GitHub repository, and full Hugging Face tree on 2026-03-26; we cannot call this date a code or weight release.
• A separate weight license and author model card on the Hub are unconfirmed; Apache-2.0 on the source code does not prove terms for each checkpoint.
• No independent replication verified on the current code exists, so reported metrics and resource numbers remain author results.

Sources¶

Agent brief¶

• Subject: project:pam, thread public-release, 1 dated events 2026-03-26 → 2026-03-26.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.