Stack_WanSoundImageToVideo

Wan 2.2-S2V-14B is Alibaba's own first-party speech-to-video model, part of the small family of task-specific models - alongside VACE and Wan Animate - that Alibaba kept shipping open under Apache 2.0 even after the numbered Wan base line went API-only past 2.2. This node, Stack_WanSoundImageToVideo, bundles its config. Worth setting expectations honestly up front: when ByteDance's rival HuMo model launched, the community's direct read was that it looked "way better than Wan S2V" specifically on facial emotion and lip-sync accuracy. S2V is the official, first-party option and a completely valid starting point, but if lip-sync fidelity is your top priority, it's worth comparing against HuMo (also in this pack, as Stack_WanHuMoImageToVideo) before committing to a workflow.

How it works. ref_image anchors identity - the subject who'll be speaking or performing. audio_encoder_output is the audio conditioning; like HuMo's node, this takes pre-encoded audio embeddings rather than a raw audio file, so you need an audio encoder node upstream to produce it. control_video is optional motion guidance if you want body movement driven by an external source rather than left to the model. ref_motion is a separate reference specifically for movement style - distinct from ref_image's identity role, this lets you decouple "what the subject looks like" from "how the subject tends to move." width, height, and length default to 832×480 at 77 frames. Output is SoundImage, a config bundle for a downstream Wan sampler.

Installing it. No extra dependency for this node in the base pack: git clone https://github.com/cardenluo/ComfyUI-Apt_Preset into custom_nodes, restart ComfyUI, or search "ComfyUI-Apt_Preset" in ComfyUI Manager. You'll need the Wan 2.2-S2V-14B checkpoint itself - ungated and Apache 2.0 on Alibaba's official HuggingFace org, so no special access request needed - plus an audio encoder node upstream to actually produce a working audio_encoder_output.

Common issues. A generation that runs fine but shows no mouth movement in sync with the audio almost always means audio_encoder_output isn't actually wired to a working encoder - it's optional at the schema level, so it's easy to leave disconnected and get a silent-seeming failure rather than an error. If lip-sync quality feels noticeably off compared to what you expected, that's a known, community-observed limitation of this specific model relative to HuMo - worth trying the HuMo node in this same pack before assuming your settings are wrong. If ref_motion doesn't seem to be changing anything, double check it's actually connected and distinct from ref_image - since both are optional image inputs, it's easy to accidentally wire the same image into both slots and get no differentiated effect between identity and motion style. And as with the other 14B-class Wan nodes in this pack, budget real VRAM for this one; it's not the lightweight option in the lineup.