model:
  fm:
    beta_dim: 10
    pose_dim: 48         # global_aa(3) + joints_aa(45), directly regressed per frame
    trans_dim: 3         # translation, directly regressed
    normalization_stats: ${oc.env:HANDFLOW_NORMALIZATION_STATS}

    condition:
      hamer_embed_dim: 1280
      hamer_n_tokens: 192       # 16×12 patches
      n_joints: 21
      skeleton_pe_freqs: 6
      cmask_ratio: 0.2

    frame_compressor:
      num_layers: 8             # cross-attention pooling layers (replaces Refiner + Compressor)

    transformer:
      hidden_size: 512
      num_heads: 16
      depth: 8                  # double-stream blocks
      depth_single_blocks: 16   # single-stream blocks
      mlp_ratio: 4.0
      qkv_bias: true
      dropout: 0.1

  hamer:
    checkpoint: ${oc.env:HAMER_CKPT}
    frozen: true