Wan2.1 VACE Reference to Video
Create videos that match the style and content of a reference image. Perfect for style-consistent video generation.
例を試す
このワークフローについて
Wan2.1 VACE Reference to Video turns a single image into a short video that preserves your reference’s style and subject while introducing motion guided by text prompts. The core of the workflow is the WanVaceToVideo node, which fuses a loaded reference image (LoadImage) with prompt text encoded by CLIPTextEncode. The diffusion backbone is provided by the Wan 2.1 VACE models—use UNETLoader and VAELoader in the “Load models here” group to select either Wan2.1-VACE-1.3B or Wan2.1-VACE-14B. A ModelSamplingSD3 node configures the sampler, and KSampler handles denoising with your chosen steps and cfg scale.
If you need stylistic adapters, LoraLoader lets you apply a matching LoRA to the same diffusion model family.
After sampling, VAEDecode reconstructs frames from latents. TrimVideoLatent helps you set exact duration, and the “First Frame control” group (ImageToMask, SolidMask, MaskToImage, RepeatImageBatch) can lock or protect parts of the initial frame for stronger identity or composition retention. PreviewImage nodes allow quick QC during iteration.
Finally, CreateVideo aggregates frames and you can export via SaveVideo (MP4) or SaveAnimatedWEBP for lightweight previews. Author guidance: default steps=20 and cfg=6.0 work well; with CausVid LoRA use fewer steps (2–4) and a low cfg (~1.0). For resolution, VACE-1.3B is suited to 480p, while VACE-14B delivers higher quality and supports 480p and 720p at the cost of longer render time.
よくある質問
Use 1.3B for faster drafts at 480p and lower VRAM. Choose 14B for better detail, stronger temporal coherence, and 720p support, noting longer render times and higher memory use.
Set the frame count in WanVaceToVideo. The helper node “Value = Length(WanVaceToVideo) - 1” keeps CreateVideo aligned. Set FPS and output settings in CreateVideo/SaveVideo. Use TrimVideoLatent to shorten or fine-tune duration without re-running the full pipeline.
Enable the “First Frame control” group. Use SolidMask or ImageToMask to define protected areas, then MaskToImage to apply them. Combine with a moderate cfg (around 6) and a fixed seed in KSampler to improve identity and layout stability.
Yes. Load a LoRA via LoraLoader that matches the exact Wan diffusion model family you’re using. Follow the author’s guidance for CausVid LoRA (steps 2–4, cfg ≈1.0). Mismatched LoRAs can degrade quality or fail to load.