Wan 2.1 Image to Video
Generate videos from images using Wan 2.1.
Try an example
About this workflow
This ComfyUI workflow turns a single image into a short video using Wan 2.1’s image-to-video pipeline. The core is the WanImageToVideo node, which prepares temporal conditioning (frame count, fps, and motion cues) and routes it into diffusion. Text guidance is encoded through CLIPTextEncode (backed by CLIPLoader loading the UMT5 XXL text encoder), while the starting image is embedded with CLIPVisionEncode using weights from CLIPVisionLoader. The model backbone is loaded via UNETLoader (Wan2.1) and VAELoader (Wan VAE).
ModelSamplingSD3 sets the correct sampling profile, and KSampler performs denoising across the sequence of latent frames. Finally, VAEDecode converts latents to RGB frames, CreateVideo assembles them at your chosen fps/size, and SaveVideo writes the result to disk.
The workflow is organized for clarity: Step1 - Load models, Step2 - Upload start_image, Step3 - Video size, Step4 - Prompt. You’ll set width/height and frame rate in the Video size group, provide a clean source image via LoadImage, and craft a text prompt (plus optional negative prompt) to steer style and content. This makes it ideal for animating character turnarounds, subtle portrait motions, or product hero shots—all from a single still—while keeping results predictable through the start image and reproducible via seed and KSampler settings.
Questions & answers
Load Wan 2.1’s UNet via UNETLoader, the matching VAE via VAELoader, the UMT5 XXL text encoder via CLIPLoader (e.g., umt5_xxl_fp8_e4m3fn_scaled.safetensors), and a CLIP vision backbone via CLIPVisionLoader. Place the files in your ComfyUI models folders (e.g., models/unet, models/vae, models/text_encoders or models/clip, models/clip_vision) or point the loader nodes directly to their locations.
Match the start image’s aspect ratio to your target width/height in the Video size group. Crop the source image accordingly before loading. Keeping the subject centered and well-lit also helps the model maintain structure during motion.
Lower the resolution (width/height), reduce the number of frames, and try fewer diffusion steps in KSampler. Start small (fewer frames, lower size), confirm the look, then scale up gradually.
Yes. The workflow encodes text with the UMT5 XXL text encoder, which works well with English and other languages (e.g., Chinese). You can also supply a negative prompt via CLIPTextEncode to suppress unwanted artifacts or styles.
For more motion, add clear action verbs and reduce overly strict descriptors that lock the pose. For smoother outputs, balance frame count and fps, keep resolution reasonable, and refine the negative prompt to remove busy backgrounds or unwanted detail that can flicker.