This ComfyUI tutorial shows you how to generate smooth, guided video clips by explicitly controlling the first and last frames using Wan 2.1 FLF2V. The core of the workflow is the WanFirstLastFrameToVideo node, which fuses three condition sources: your prompt embeddings from CLIPTextEncode, visual guidance from CLIPVisionEncode of the start/end images, and the Wan 2.1 UNet loaded by UNETLoader. The workflow is configured for 720p output and 16 frames (F16), giving you a compact, visually coherent clip that cleanly transitions from the first frame to the last.

Under the hood, the model components are loaded in the "Load Models Here" group: VAELoader and VAEDecode handle latent-to-RGB decoding of frames; CLIPLoader and CLIPTextEncode create text and (optionally) negative prompt embeddings; CLIPVisionLoader with CLIPVisionEncode extract visual features from your start and end images. ModelSamplingSD3 defines the timestep/schedule used by KSampler to drive diffusion steps for the Wan 2.1 UNet. Finally, CreateVideo assembles the decoded frames and SaveVideo writes the MP4 to disk. This setup is ideal when you want precise control over motion and composition—e.g., matching storyboard keyframes—while still benefiting from text-driven styling.