This workflow turns a pair of images into a short, coherent video by leveraging Wan 2.1’s inpainting-based video diffusion. You provide a start frame and an end frame; the WanFunInpaintToVideo node then guides the model to synthesize the in‑between motion while preserving subject and scene details anchored in your images. Text prompts (via CLIPTextEncode) steer style and content, while negative prompts help suppress artifacts.

Under the hood, Step1 loads the core components: UNETLoader (Wan2.1 / Wan), VAELoader/VAEDecode for latent <-> pixel space conversion, and CLIPLoader for text conditioning. Your start/end frames are brought in with LoadImage and encoded using CLIPVisionLoader + CLIPVisionEncode to produce image conditioning for the UNet. KSampler drives the diffusion process using ModelSamplingSD3 for the correct sigma schedule, with CFGZeroStar and SkipLayerGuidanceDiT stabilizing guidance. The Attention Booster group applies UNetTemporalAttentionMultiply to strengthen temporal consistency across frames. Finally, CreateVideo assembles frames and SaveVideo writes the finished clip.