configs/inference.yaml

# 预训练模型路径
dtype: "bf16"
text_encoder_path: pretrained_models/Wan2.1-T2V-14B/models_t5_umt5-xxl-enc-bf16.pth
image_encoder_path: None
dit_path: pretrained_models/Wan2.1-T2V-14B/diffusion_pytorch_model-00001-of-00006.safetensors,pretrained_models/Wan2.1-T2V-14B/diffusion_pytorch_model-00002-of-00006.safetensors,pretrained_models/Wan2.1-T2V-14B/diffusion_pytorch_model-00003-of-00006.safetensors,pretrained_models/Wan2.1-T2V-14B/diffusion_pytorch_model-00004-of-00006.safetensors,pretrained_models/Wan2.1-T2V-14B/diffusion_pytorch_model-00005-of-00006.safetensors,pretrained_models/Wan2.1-T2V-14B/diffusion_pytorch_model-00006-of-00006.safetensors
vae_path: pretrained_models/Wan2.1-T2V-14B/Wan2.1_VAE.pth
wav2vec_path: pretrained_models/wav2vec2-base-960h
exp_path: pretrained_models/OmniAvatar-14B
num_persistent_param_in_dit:  # You can set `num_persistent_param_in_dit` to a small number to reduce VRAM required. 

reload_cfg: True
sp_size: 1

# 数据参数
seed: 42
image_sizes_720: [[400, 720], 
             [720, 720], 
             [720, 400]]
image_sizes_1280: [
             [720, 720], 
             [528, 960],
             [960, 528],
             [720, 1280],
             [1280, 720]]
max_hw: 720 # 720: 480p; 1280: 720p
max_tokens: 30000  
seq_len: 200
overlap_frame: 13 # must be 1 + 4*n
guidance_scale: 4.5
audio_scale: 
num_steps: 16
fps: 25
sample_rate: 16000
negative_prompt: "Vivid color tones, background/camera moving quickly, screen switching, subtitles and special effects, mutation, overexposed, static, blurred details, subtitles, style, work, painting, image, still, overall grayish, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn face, deformed, disfigured, malformed limbs, fingers merging, motionless image, chaotic background, three legs, crowded background with many people, walking backward"
silence_duration_s: 0.3
use_fsdp: False
tea_cache_l1_thresh: 0 # 0.14 The larger this value is, the faster the speed, but the worse the visual quality. TODO check value