WanSoundImageToVideoExtend

WanSoundImageToVideo gives you one clip. WanSoundImageToVideoExtend gives you the next one - and the next, and the next. It's the continuation half of Wan 2.2 S2V: you feed in the latent of the video you just made, tell it how long the new chunk is, and it keeps the same character and, critically, continues the audio where the last chunk left off. Same conditioning node family, same KSampler at the end, one extra input. It shipped a day after the base S2V node, in August 2025, when Alibaba's "minute-level generation" promise ran straight into the 81-frame Wan-native clip limit.

How it works

The node takes your previous chunk's video_latent and does three things with it. It infers the width, height, and batch size from the latent's shape (so you can't silently change resolution mid-chain). It works out how far into the audio the previous chunk got - each latent frame represents four video frames - and starts the new audio window there. And it uses the last 19 latent frames of the previous chunk as a reference motion condition, which is what keeps the character's motion flowing across the cut instead of restarting. It then builds a fresh zero latent for the new chunk and attaches the continuing audio embedding.

The inputs

• positive / negative - text conditioning, same as the base node.
• vae - same Wan 2.1 VAE as WanSoundImageToVideo.
• video_latent - required; the latent output of a previous S2V node (either base or extend).
• length - frames for the new chunk (default 77).
• audio_encoder_output, ref_image, control_video - optional, same roles as the base node.

Outputs: positive, negative, latent → KSampler. The output latent is the new chunk only - you decode each chunk separately and concatenate the videos, or wrap the chain in a video-join node.

Getting it

Ships with ComfyUI core. Same model files as WanSoundImageToVideo: the S2V checkpoint, the wav2vec2 audio encoder in models/audio_encoders, wan_2.1_vae, umt5.

Where people get burned

Extending past the audio length. If your combined chunks outlast the audio, the tail gets zero audio conditioning - this is the exact bug the ComfyUI team fixed in #9608 shortly after release, so it's a smooth-ish failure now, but you're still generating dead-air footage. Keep chunks inside the audio, or loop/extend the audio track. Chunk consistency. Resolution and batch come from the latent, so keep every chunk identical or the chain visibly breaks. And expect some identity drift across cuts - it's much better than a cold restart, but it's not seamless. If you want truly invisible long-form joins on Wan, the community's answer is SVI's error-recycling pipeline, which treats this node family as one tool among several.