# Sanitized public config for pancx/StreamAvatar-AROD.
# Runtime paths are relative to the StreamAvatar repository root.

seed: 123

checkpoint_path: checkpoints/last.ckpt
wav2vec_path: pretrained_model/wav2vec2-base-960h
output_dir: outputs/blockwise_stream_distill_cross_fm_teacher_cache_anchor_pretrain_60k

model:
  module_name: model.motion_generation.motion_gen_gpt_flowmatching_addaudio_linear_twowavencoder
  class_name: Audio2FaceGPT
  pose_fps: 25
  audio_sr: 16000
  audio_fps: 16000
  wav2vec_layer: 8
  cbh_window_length: 96
  vae_codebook_size: 512
  seed: 123
  prev_audio_frames: 32
  ema_decay: 0.999
  cfg_audio: 0.5
  cfg_audio_other: 0.5
  cfg_anchor: 0
  cfg_all: 1.0
  cfg_fusion: 2.0
  cfg_audio_anchor: 1.0
  drop_audio: 0.1
  drop_audio_other: 0.1
  drop_anchor: 0.1
  drop_gpt: 0.1
  drop_style: 0.3
  drop_other: 0.0
  drop_self: 0.0
  drop_fusion: 0.1
  drop_prev_motion: 0.0
  cfg_gpt: 1.0
  cfg_style: 1.0
  cfg_prev_motion: 0.0
  sync_only: true
  use_init_motion: false
  strength: 1.0
  window_size: 8
  total_iter: 600000
  eval_metrics: false
  version: '0506'
  two_step: false
  infer_32: false
  qk_norm: rms_norm_across_heads
  eps: 1.0e-06
  cross_attn_norm: true
  added_kv_proj_dim: null

teacher:
  denoising_steps: 1
  use_ema: true
  guidance_mode: all_only

student:
  architecture: cross_fm
  block_frames: 2
  history_frames: 32
  hidden_dim: 512
  layers: 6
  heads: 8
  dropout: 0.1
  detach_between_blocks: true

loss:
  velocity_weight: 1.5
  acceleration_weight: 0.2
  boundary_weight: 0.3
  delta_weight: 0.0

noise_scheduler:
  num_train_timesteps: 1000
  shift: 1
  use_karras_sigmas: false