# Passed directly to WanDiffusionWrapper(**model_kwargs).
model_kwargs:
  model_name: Wan2.2-TI2V-5B
  timestep_shift: 5.0
  num_frame_per_block: 8
  local_attn_size: 32

use_ema: false
output_folder: videos/longlive2
num_samples: 1
save_latents_only: false

data:
  data_path: /path/to/longlive2/inference_prompts
  # Latent tensor shape [B, F, C, H, W]. F is also the default number of
  # generated latent frames unless num_output_frames is provided.
  image_or_video_shape:
  - 1
  - 128
  - 48
  - 44
  - 80

inference:
  sampling_steps: 4
  sink_size: 8
  guidance_scale: 1.0
  multi_shot_sink: true
  multi_shot_rope_offset: 8
  streaming_vae: false
  async_vae: false
  vae_type: wan # wan, mg_lightvae, mg_lightvae_v2
  vae_device: "cuda:2"

checkpoints:
  generator_ckpt: /path/to/longlive2/generator_checkpoint.pt
  lora_ckpt: /path/to/longlive2/lora_checkpoint.pt

# Presence of this section enables LoRA inference; remove it for full-generator inference.
adapter:
  type: lora
  rank: 128
  alpha: 128
  dropout: 0.0
  verbose: true

logging:
  seed: 0