Talking video

A face, a voice sample and a script — and the face says it.

What it takes

• A picture of the face
• A few seconds of clean speech as WAV — what the voice SOUNDS like, not what it says
• The line you want spoken

What it gives back

• An MP4 with the audio muxed in
• The generated read as its own WAV, so you can reuse it

What actually runs

• F5-TTS clones the sample and reads your script in that voice. It is zero-shot: a few seconds is the whole model of the voice.
• wav2vec2 turns that read into the conditioning that drives the mouth.
• Wan 2.2 S2V animates the face against it. The clip is as long as the read, so it never stops mid-word.

Worth knowing. Leave the script empty and the uploaded audio is used exactly as recorded — useful when you already have the take you want.

Running it

The .json attached to this listing is the graph itself — the one the Forge runs, not a tidied-up copy. Values in {{double braces}} are the placeholders filled in per job; replace them with real values and it runs in any ComfyUI new enough for the nodes it uses.

Full setup — the model manifest, the directories, the traps — is at /workflows/talk.

Licence and credit

The graph is ours and you may use it for anything, including commercially. The models it loads are not ours: each carries its own licence from whoever trained it, and that is the one that governs what you may do with the output.