---
license: other
license_name: markermogen-research-only
license_link: https://huggingface.co/CurMack/markermogen-paper512/blob/main/LICENSE.txt
tags:
- human-motion
- smplx
- diffusion
- experimental
---

## Current matching source and checkpoints:2026-10-20-r1

[Release files](releases/20261020/README.md) · [Chinese inference guide](https://github.com/CurMack/markermogen/blob/body-reconstruction-20261020-r1/docs/RECONSTRUCTION_20261020_ZH.md). The r1 patch preserves existing inference outputs and every guidance trace event. Weight bytes are unchanged from the fully verified20261020 release; experimental quality limitations remain.


## 2026-10-20 matching body reconstruction release

[Checkpoints and manifest](releases/20261020/README.md) · [source and Chinese inference guide](https://github.com/CurMack/markermogen/blob/32cd702f0d4b864df91192f4695e0e3f44717c80/docs/RECONSTRUCTION_20261020_ZH.md). Default: learned observed prior NORMAL12000, unguided. The latest joint diffusion NORMAL1370 is an experimental alternative; paired TRAIN temporal/hand regressions prevent recommending it. Full TEST/OOD visual quality remains unaccepted. Original models and histories below are retained.

# MarkerMoGen native paper512 experimental checkpoints

Source: https://github.com/CurMack/markermogen/tree/dc024c28dc1a12555841ef99165504791def332e

Own inference-only 306D body/finger x0 priors with normalized clean training statistics and globally SHA-bound development profiles. These checkpoints load strictly and have finite weights, but reconstruction quality is not accepted and superiority to MAMMA/SAM3DBody is not established. The current lowvis50 checkpoint is early step1000; the separate AMASS-only prior finished100000 steps. Do not initialize fresh mixed training from the old AMASS prior.

Public third-party mirrors retain their original licenses. MammaNet, its marker map, SMPL and SMPL-X must be obtained from the official sites under their licenses; they are not mirrored here. See assets.json for complete availability. No raw motion/video, SMPL-X buffers, actor information, secrets or signed release links are included.

Use the GitHub current guide and requirements-paper512.txt, obtain separately authorized decoder assets, then run markermogen.paper512.infer with the exported model.pt and its global_development_profile.json. All sequences for a candidate share one checkpoint and profile. Do not use these development profiles to open final held-out sequences.

## Repaired CUDA inference source

Use source commit https://github.com/CurMack/markermogen/tree/330ce8bdc124bd4df6a9acf714872087785fac05 and the matching `global_development_cpu_geometry_profile.json` next to each model. Triangulation uses identical CPU equations for all inputs; root/shape fitting, diffusion and guidance run on GPU. All 56 geometry checks were bitwise equal to the original CPU equations, and CUDA validation now runs. This is a runtime correction; quality remains experimental. Initial profiles and their source commit are retained for reproducibility.

ViTPose, SAM3DBody and MHR are available with their licenses and exact official revisions in assets.json and official_sources.json. MammaNet and SMPL/SMPL-X require official authorized downloads.

## Repaired CUDA inference source

Use source commit https://github.com/CurMack/markermogen/tree/330ce8bdc124bd4df6a9acf714872087785fac05 and the matching `global_development_cpu_geometry_profile.json` next to each model. Triangulation uses identical CPU equations for all inputs; root/shape fitting, diffusion and guidance run on GPU. All 56 geometry checks were bitwise equal to the original CPU equations, and CUDA validation now runs. This is a runtime correction; quality remains experimental. Initial profiles and their source commit are retained for reproducibility.

ViTPose, SAM3DBody and MHR are available with their licenses and exact official revisions in assets.json and official_sources.json. MammaNet and SMPL/SMPL-X require official authorized downloads.