| import torch | |
| from diffsynth import save_video | |
| 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"), | |
| ], | |
| ) | |
| pipe.enable_vram_management() | |
| # Text-to-video | |
| 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, "video1.mp4", fps=15, quality=5) | |