| 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="Wan-AI/Wan2.1-T2V-14B", origin_file_pattern="diffusion_pytorch_model*.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/Wan2.1-T2V-14B_full/epoch-1.safetensors") |
| pipe.dit.load_state_dict(state_dict) |
| pipe.enable_vram_management() |
|
|
| video = pipe( |
| prompt="from sunset to night, a small town, light, house, river", |
| 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", |
| seed=1, tiled=True |
| ) |
| save_video(video, "video_Wan2.1-T2V-14B.mp4", fps=15, quality=5) |
|
|
