Wan22 Sound Image To Video Extend (Tiled VAE Encode)

Five seconds of talking-head video isn't enough for most projects, and that's what this node exists for. It's the extension companion to the tiled S2V node: feed it the video_latent from a completed S2V clip, hand it more audio, and it conditions a continuation that keeps the same person, the same motion style, and the same sync. Everything gets tiled, so chaining a long speech doesn't have to chain OOM errors instead.

What it actually does

Where the plain S2V node takes a ref_image and builds its conditioning from scratch, this one takes a video_latent - the latent your previous S2V generation produced - and treats it as the reference motion. The node reads the latent's shape to figure out width, height, batch_size, and where in the audio stream to start: the frame offset is computed as the number of latent frames × 4, so the new clip's audio sync begins right where the old one ended.

That continuation latent is fed in as reference_motion (the last 19 latent frames are kept), so the model knows exactly where the last clip left off and what motion to extend. Then the new length worth of frames is generated on top. You can still pass a ref_image for identity anchoring and a control_video for pose driving.

Inputs and outputs

Notice what's missing: no width / height / batch_size - those are derived from video_latent. Required: positive / negative conditioning, vae, video_latent, length, and the tiling quartet. Optional: audio_encoder_output, ref_image, control_video.

Outputs: positive, negative, latent. The new latent is the continuation; you can chain this node to itself - feed its latent back into video_latent and extend again.

Installing it

ComfyUI Manager → search Fossiel or WAN-Additional-Tilers, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Fossiel/ComfyUI-Fossiel-WAN-Additional-Tilers

Restart ComfyUI. No Python deps, no bundled models - your own S2V weights, CLIP, VAE, and audio encoder.

Gotchas

Continuity is the whole game here, and it's the same weakness the Wan ecosystem has always had: identity and appearance can drift across chunk boundaries, so a long multi-extend chain wants a ref_image on each segment to keep the face locked. Audio sync is computed per segment too, so keep each length close to the model's native range rather than trying one giant clip. And the pack's standing rule: keep ComfyUI current for the temporal tiling arguments - an old build will reject encode_tiled() outright.