Marionette / config.json
SII-Sarfflow's picture
Add files using upload-large-folder tool
bdb49ef verified
Raw
History Blame Contribute Delete
2.03 kB
{
"_comment": "Uploaded to the Hub as config.json. It describes the three released checkpoints and the state they operate on; it is NOT a transformers or diffusers config and deliberately carries no model_type / architectures / _class_name, because nothing in this repository is loadable by those libraries. Loading is done by the code at https://github.com/AlayaLab/Marionette (fetch_weights.sh, run_demo.sh). It also gives the Hub a file to count downloads against -- see hf/DOWNLOAD_COUNTING.md.",
"name": "Marionette",
"paper": "https://arxiv.org/abs/2608.14530",
"code": "https://github.com/AlayaLab/Marionette",
"project_page": "https://alayalab.github.io/Marionette/",
"license": "see LICENSE.assets in the code repository: non-commercial research use only",
"pipeline": ["dynamics", "bridge", "observation"],
"world_state": {
"dim": 276,
"entities": 2,
"fps": 20,
"contents": "per-entity articulated skeletons, metric root trajectories, rotations"
},
"stages": {
"dynamics": {
"weights": ["dynamics/action_gpt.pt", "dynamics/pose_gpt.pt"],
"role": "autoregressive prediction of the 276-dimensional world state from a seed",
"components": ["ActionGPT", "PoseGPT"]
},
"bridge": {
"weights": [],
"role": "closed-form world-space geometry and occlusion; zero learnable parameters"
},
"observation": {
"weights": ["observation/diffusion_pytorch_model.safetensors"],
"role": "control-conditioned video diffusion; paints appearance onto the rendered geometry",
"base_model": "alibaba-pai/Wan2.2-Fun-5B-Control",
"base_model_distributed_here": false,
"resolution": [704, 1280],
"chunk_frames": 81,
"rollout": "chunk-relay autoregressive"
}
},
"scope": {
"dynamics": "single monster (em19), one stage, one weapon type",
"observation": "26 monsters",
"note": "the two stages were trained on different slices of the same corpus; end-to-end runs are limited by the narrower one"
}
}