How is this different from optical-flow interpolation?

Rather than warping pixels between frames, WanVaceToVideo synthesizes intermediate content in latent space using the Wan2.1 VACE prior. This yields more stable identity, fewer artifacts on large disocclusions, and better adherence to your text prompt or LoRA style.

Why must (Length − 1) be divisible by 4 and the size be divisible by 16?

The VACE transition and UNet operate on stride-aligned temporal and spatial blocks. Using a 4n+1 frame count (e.g., 5, 9, 13…) and dimensions divisible by 16 ensures the sampler and decoder align correctly, preventing artifacts and dropped frames.

When should I use 1.3B vs 14B, and what resolutions are supported?

VACE-1.3B is optimized for 480p and is lighter to run. VACE-14B supports both 480p and 720p and typically delivers stronger detail and motion coherence. Choose 14B for higher fidelity or 720p; use 1.3B when resources are limited or for fast previews.

How do LoRAs affect quality and speed in this workflow?

With LoraLoader enabled, you can guide style or motion while reducing steps. For a CausVid-style LoRA, set steps to 2–4 and cfg ≈ 1.0; otherwise, defaults like 20 steps and cfg ≈ 6.0 work well. Start LoRA strength at 0.3–0.7 to balance guidance and temporal stability.

Wan2.1 VACE First-Last Frame - ComfyUI Workflow

Wan2.1 VACE First–Last Frame introduces endpoint-anchored video synthesis: you provide a starting frame and an ending frame, and the model plans a temporally coherent path between them. Instead of naive optical-flow interpolation, the workflow synthesizes intermediate content in latent space using Wan2.1’s VACE prior, guided by your prompt and optional LoRA. This delivers stable identity, style continuity, and believable motion planning across the entire sequence.

In ComfyUI, the pipeline pairs WanVaceToVideo with a ModelSamplingSD3 profile, KSampler, and a strict frame/layout regimen: spatial dimensions must be divisible by 16, and the total frame count follows a 4n+1 rule (Length − 1 divisible by 4). This structure, plus TrimVideoLatent for exact clip length, produces transitions that are both smooth and controllable. The 14B model supports clean 480p and 720p outputs, while the lighter 1.3B is optimized for 480p. With optional LoRA (e.g., CausVid), you can cut sampling steps dramatically while maintaining temporal coherence.

FAQ

Часто Задаваемые Вопросы