name: dit3d_pose
variant: full
pos_emb_type: rope_3d
patch_size: 2
hidden_size: 384
depth: 12
num_heads: 6
mlp_ratio: 4.0
use_gradient_checkpointing: False
conditioning:
  modeling: film # concat (concatenate conditioning to the video input across channels), film (FiLM conditioning)
  type: ${algorithm.camera_pose_conditioning.type}
  dim: null # this will be overwritten at dfot_video_pose.py
external_cond_dropout: 0.1 # probability of dropping a camera pose of each video during training