MiniMax H3 Image + Audio-to-Video Lip-Sync Demo (with Irodori-TTS v3)

Melvin Vivas · X video post · 2026-08-07 · 0:10 · 228 views · Open on X

Topics: AI Dev Tools & Productivity, Industry Trends & Job Market · Level: beginner

Summary

This is a 10-second post by Melvin Vivas. It shares a demo of MiniMax H3 from kohya_tech's quoted post. The video was made from a starting image plus an audio track that Irodori-TTS v3 generated. MiniMax H3 then produced only the video. Lip-sync matches the audio well, but large body movements still look a bit wrong. The only spoken line is "いい仕事してますね" ("Nice work").

Key points

• Pipeline: generate speech with a TTS model (Irodori-TTS v3), then give MiniMax H3 a starting image plus that audio to generate the video.
• MiniMax H3 only generates the video. The audio comes from the separate TTS step.
• Lip-sync matched the supplied audio well in this test.
• Limitation: large movements in the clip still look somewhat off.
• The original experiment was posted by kohya_tech (Kohya). Melvin Vivas reshared it as a quote post, so the post has no tutorial or code.

Resources mentioned

• MiniMax H3 · tool · huggingface.co · free
  MiniMax video generation model that turns a starting image and an audio track into a lip-synced video.
  Also in: Generating AI Video Locally with MiniMax H3 in ComfyUI on an RTX 3090 (Melvin Vivas on X · notes), Running MiniMax H3 Locally on a Single RTX 3090 (Demo) (Melvin Vivas on X · notes)
• Irodori-TTS v3 · tool · github.com · free
  Text-to-speech model, used here to make the audio track that drives the video generation.
• kohya_tech (Kohya) · person · x.com · free
  Generative-AI developer on X who ran the original MiniMax H3 lip-sync experiment.

Try this

• Build a talking-avatar pipeline: generate speech with a TTS model, then feed a portrait image plus that audio to an image + audio-to-video model such as MiniMax H3, and check lip-sync quality against how well it handles large movements.