{
  "_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"
  }
}