Wan2.1 VACE Control Video
Generate videos by controlling input videos and reference images using Wan VACE.
试试一个示例
关于此工作流
Wan2.1 VACE Control Video is a guided video-to-video workflow that preserves the motion and timing of a source video while restyling or re-characterizing it with a text prompt and a reference image. It loads Wan2.1 VACE checkpoints (1.3B or 14B) via UNETLoader and VAELoader, encodes your text with CLIPTextEncode, and conditions generation with control signals derived from simple image preprocessing (Canny by default). GetVideoComponents reads frames and FPS from the input clip, and TrimVideoLatent keeps the latent sequence aligned with your chosen frame range and resolution before sampling.
Under the hood, ModelSamplingSD3 configures the expected sampler schedule for Wan2.1 VACE, and KSampler performs denoising on a sequence of video latents assembled by WanVaceToVideo using your control video, prompt, and reference image. VAEDecode converts the final latents back to frames, CreateVideo assembles them at the desired FPS, and you can export either MP4 (SaveVideo) or animated WebP (SaveAnimatedWEBP). An optional CausVid LoRA (LoraLoader) can accelerate or stabilize motion with very low steps, though you may need to tune its strength.
The 14B model yields higher fidelity (and supports 720p) but runs slower and uses more VRAM; the 1.3B variant is faster and suited to 480p outputs.
常见问答
Use 1.3B for faster iteration and 480p outputs. Choose 14B if you need higher fidelity or 720p. The 14B model is slower and requires more VRAM but typically yields cleaner detail and temporal stability.
Lower the LoRA strength toward 0.3–0.5, slightly increase steps (e.g., from 2 to 3–4), and keep CFG near 1.0 for the LoRA preset. If artifacts persist, bypass the LoRA entirely or relax the Canny thresholds so edges are less constraining.
TrimVideoLatent lets you limit or align the frame range. GetVideoComponents reads FPS from the source; you can keep it or override during CreateVideo. Shorter clips and moderate FPS will reduce VRAM and speed up sampling.
Yes. Replace the Canny node with another comfyui_controlnet_aux preprocessor (e.g., Lineart or Depth) and route its output to the same control input used by WanVaceToVideo. Different preprocessors change how strictly structure is preserved.