|
|
import torch |
|
|
from PIL import Image |
|
|
from diffsynth import save_video, VideoData, load_state_dict |
|
|
from diffsynth.pipelines.wan_video_new import WanVideoPipeline, ModelConfig |
|
|
|
|
|
|
|
|
pipe = WanVideoPipeline.from_pretrained( |
|
|
torch_dtype=torch.bfloat16, |
|
|
device="cuda", |
|
|
model_configs=[ |
|
|
ModelConfig(model_id="krea/krea-realtime-video", origin_file_pattern="krea-realtime-video-14b.safetensors", offload_device="cpu"), |
|
|
ModelConfig(model_id="Wan-AI/Wan2.1-T2V-14B", origin_file_pattern="models_t5_umt5-xxl-enc-bf16.pth", offload_device="cpu"), |
|
|
ModelConfig(model_id="Wan-AI/Wan2.1-T2V-14B", origin_file_pattern="Wan2.1_VAE.pth", offload_device="cpu"), |
|
|
], |
|
|
) |
|
|
state_dict = load_state_dict("models/train/krea-realtime-video_full/epoch-1.safetensors") |
|
|
pipe.dit.load_state_dict(state_dict) |
|
|
pipe.enable_vram_management() |
|
|
|
|
|
|
|
|
video = pipe( |
|
|
prompt="a cat sitting on a boat", |
|
|
num_inference_steps=6, num_frames=81, |
|
|
seed=0, tiled=True, |
|
|
cfg_scale=1, |
|
|
sigma_shift=20, |
|
|
) |
|
|
save_video(video, "output.mp4", fps=15, quality=5) |
|
|
|
