ode_config.yaml

trainer: ode
model_name: causal_wan

generator_grad:
  model: true

denoising_step_list:
- 1000
- 750
- 500
- 250

generator_task: causal_video
generator_fsdp_wrap_strategy: size
text_encoder_fsdp_wrap_strategy: size
mixed_precision: true
seed: 0

wandb_host: https://api.wandb.ai
wandb_key: 
wandb_entity: 
wandb_project:
wandb_name: 

sharding_strategy: hybrid_full
lr: 2.0e-06
beta1: 0.9
beta2: 0.999

data_path: ../lmdb/vidprom_ode_lmdb
batch_size: 8  # Per GPU
total_batch_size: 64  # 8 GPUs x 8 batch_size
log_iters: 250  # Save checkpoint every 250 iterations
max_iters: 3000  # Stop training after reaching this many iterations (will save at 250, 500, 750, 1000) 
output_path: 

distribution_loss: ode
gradient_checkpointing: true
num_frame_per_block: 3
warp_denoising_step: true

# Video generation shape for inference
image_or_video_shape:
- 1
- 21
- 16
- 60
- 104

# Model kwargs - must match data generation settings
model_kwargs:
  timestep_shift: 5.0

# Optional negative prompt (Chinese + English)
negative_prompt: '色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走'