This ComfyUI workflow turns a single reference image and an audio track into a synchronized video using the Wan 2.2 S2V (sound-to-video) model. It loads the Wan backbone with UNETLoader, CLIPLoader, and VAELoader, and sets the correct schedule via ModelSamplingSD3. Your text prompt is encoded by CLIPTextEncode, while the audio is processed by AudioEncoderLoader and AudioEncoderEncode. These signals, together with the image from LoadImage, drive WanSoundImageToVideo to synthesize motion that follows the rhythm and phrasing of the audio. The KSampler controls steps, CFG, and seed; VAEDecode converts latents to frames; and CreateVideo assembles the final clip.

Two engineering details make the workflow practical for production: a first-frame stabilization and scalable length. The Fix overbaked first frame group uses LatentCut and LatentConcat to duplicate the initial latent frame before decoding, then discards the first decoded frame to avoid the common VAE “overbaked” look on frame 0. For long clips, the Video S2V Extend subgraphs let you chain segments in chunks (77 frames by default) using LatentCut/LatentConcat, so you can build minute-level videos without blowing up VRAM. The workflow also supports a lightning path: with a Wan 2.2 lightning LoRA you can sample in 4 steps (recommended CFG ~1.0). If you prefer quality and stability, use the standard path at ~20 steps (CFG ~6.0).