Wan 2.2 14B First-Last Frame to Video
Generate smooth video transitions by defining start and end frames.
Try an example
About this workflow
This workflow introduces First–Last Frame to Video (FLF2V) for Wan 2.2 (14B), a conditioning method that lets you specify both the opening and closing frames and have the model synthesize a smooth, prompt-guided journey between them. Instead of expanding from a single keyframe or relying on optical-flow interpolation, the WanFirstLastFrameToVideo node enforces boundary conditions at t=0 and t=1 inside the diffusion process. That means the start and end images are respected as ground truths while the model generates new, coherent in-betweens that follow your text prompt.
Under the hood, the workflow loads Wan 2.2 with UNETLoader, a compatible text encoder via CLIPLoader (UMT5 XXL FP8 scaled), and VAE components, then runs KSamplerAdvanced with an SD3-style schedule (ModelSamplingSD3). An optional “lightning” 4-step LoRA can dramatically speed up renders at a modest quality trade-off. The result is a practical, reproducible pipeline in ComfyUI that improves on prior image-to-video approaches by jointly conditioning on both endpoints, reducing identity drift, preserving scene intent, and producing transitions that feel authored rather than warped.
Questions & answers
FLF2V is generative. It conditions the diffusion process on both the start and end images concurrently, so the model synthesizes new, coherent midframes rather than warping pixels between frames. Compared with single-frame I2V, it reduces identity drift and guarantees the final frame matches your target image.
The fp8_scaled configuration significantly lowers VRAM use. As a reference from the author’s notes, 640×640 uses about 84% VRAM on an RTX 4090D 24GB. Enabling the 4-steps lightning LoRA can speed up renders further, but expect some quality and temporal consistency trade-offs.
Yes. The workflow uses CLIPLoader with a compatible text encoder (e.g., UMT5 XXL FP8 scaled) and sets ModelSamplingSD3 to match Wan 2.2’s expected schedule. KSamplerAdvanced then handles steps, CFG, and seed. Using the provided components as wired ensures stable, reproducible results.
Use start/end images with matching aspect ratios and related content, keep the prompt intent aligned with both frames, and avoid conflicting style cues. Increase steps or CFG for sharper adherence, and reduce resolution or total frames if you encounter VRAM pressure.