Wan2.2-S2V is documented as an audio-driven video-generation pathway that combines a portrait with audio to animate speaking or performance.
A portrait is not enough
It is distinct from the general Wan2.2 text-to-video and image-to-video workflows: the sound track influences visible actions and mouth movement. The official ComfyUI guide provides 14B model variants, including precision choices, and describes extension-based long-form possibilities. Those are workflow features, not proof that a ten-minute monologue will remain visually consistent.
Prepare input audio before generation
Begin with a clean, licensed voice track and a script that has been edited for facts and spoken pronunciation. Remove clipping and unwanted background noise in a dedicated audio editor without stripping expressive timing. Use a legally cleared starting image that depicts a willing human participant or a fictional character. Do not animate a third party’s recognizable face simply because it is technically possible. Capture the exact speaker approval and intended publication context.
Set up a model-specific graph
Use the current ComfyUI native workflow rather than dropping S2V weights into an unrelated Wan2.2 graph. Note the checkpoint precision, model hash, audio samplerate requirements described by the template, encoder files and extension strategy. If memory is tight, prefer an officially documented precision alternative and verify visual quality; do not invent a VRAM guarantee without an identified GPU and tested workflow. Keep all external weight licenses and custom-node revisions with the project file.
Segment by editorial meaning
For a 90-second speech, divide the track into coherent clauses or scenes with breath points. Review short approved segments before extending, because errors in a speaker’s face or pose may compound over time. Maintain a master timecoded script, and align subtitle timing to the actual final audio rather than to the source transcript alone. Check mouth closures, sibilants and interruptions at normal playback speed; render quality is not determined by one attractive freeze-frame.
Where traditional compositing wins
A frozen portrait, interview subtitles and B-roll may explain a topic more accurately than continuously moving a synthetic face. Use ordinary cuts when a character changes outfit or a documented real person should not appear to make an invented statement. Record whether the visual was generated, who consented to its creation and which model version was involved. Rights to music and images stay separate from rights to use a model’s checkpoint.
Relevant alternatives and trade-offs
Hedra Character 3 has an explicit hosted API audio-duration specification; LTX-2.5 is a broader audio-video model rather than a dedicated image-plus-speech portrait workflow.