Why must width/height be divisible by 16 and length-1 be divisible by 4?

WanVaceToVideo and the VAE downsample spatially by powers of two, so dimensions must align to 16 for clean latent shapes. The temporal pathway expects sequence lengths that satisfy (frames - 1) % 4 == 0. TrimVideoLatent enforces the temporal rule; set your target resolution to multiples of 16 for the spatial rule.

Which model should I use: Wan2.1-VACE-1.3B or 14B?

Use 1.3B for faster runs and 480p outputs on modest GPUs. Choose 14B for higher quality and, if your VRAM allows, higher resolutions such as 720p. The 14B model produces better detail but is slower and more memory-intensive.

How do I control where outpainting happens?

Use ImagePadForOutpaint to add canvas on chosen sides (top/bottom/left/right). Then create a mask that protects the original video region (ImageToMask/MaskToImage) and replicate it across frames with RepeatImageBatch. The model synthesizes only in the unmasked, newly padded areas.

My output is black/misaligned. What should I check?

Verify three points: 1) Output width/height are divisible by 16, 2) TrimVideoLatent is applied so (frames - 1) is divisible by 4, 3) VAEDecode is connected before CreateVideo/SaveVideo. Also confirm fps and frame count from GetVideoComponents, and that your mask batch length matches the video length.

Wan2.1 VACE Outpainting - ComfyUI Workflow

This ComfyUI workflow teaches you how to outpaint a video—expanding its canvas while keeping the original content intact—using Wan 2.1 VACE. It loads your source clip with LoadVideo and extracts dimensions and frame count via GetVideoComponents and PreviewImage. You then define where to expand using ImagePadForOutpaint and a mask built from ImageToMask/MaskToImage, replicated across time with RepeatImageBatch. Prompts are encoded with CLIPTextEncode, and generation runs through KSampler with ModelSamplingSD3 scheduling on a Wan 2.1 VACE UNet/Decoder stack (UNETLoader, VAELoader), with VAEDecode converting latents to frames. WanVaceToVideo handles video-specific conditioning, and TrimVideoLatent ensures temporal shapes align. The result is assembled with CreateVideo and saved via SaveVideo.

Technically, two alignment rules matter for stable inference: the spatial size you set for outpainting must be divisible by 16 (a constraint of WanVaceToVideo and the VAE), and the sequence length must satisfy the temporal constraint (Length - 1 divisible by 4), which TrimVideoLatent addresses. Because core Comfy nodes don’t yet expose video size/frame-count utilities, this template uses PreviewImage as a practical workaround; you can swap in custom utility nodes if available. The workflow includes model groups for Wan2.1-VACE-1.3B (lighter, recommended for 480p) and Wan2.1-VACE-14B (higher quality, suitable for higher resolutions like 720p if resources allow). Default generation settings are steps: 20 and cfg: 6.0, with optional CausVid LoRA tuning at steps: 2–4 and cfg: 1.0.

FAQ

Frequently Asked Questions