Introduction

You've got a great AI shot, but one word in the dialogue is wrong. Do you re-roll the whole clip? H3 Audio Retake Studio says no. It's a new open-source, browser-based studio from developer Nebuluss that lets you select a passage on a waveform, make a new take of the line, and re-sync the speaker's mouth with MiniMax H3.

It runs as a ComfyUI custom node and opens at http://127.0.0.1:8188/audio-retake/, so there's no separate server. The interface drives the ComfyUI jobs for you, so you don't need to wire a separate graph for every fix. The repo went up on October 31, picked up joint retakes and project tools the same morning, and the author shared a narrated 2½-minute demo. The app code is MIT.

What you can do with it

Pick the exact words

  • Drag over a passage on the waveform, zoom up to 32×, and type IN and OUT points to the millisecond
  • Nudge the edges by 1 ms, 10 ms, one frame or 100 ms, or use Snap edges to pauses
  • Transcribe builds a clickable word list with a local Whisper model. Click a word to select it

Make a new take

  • Type the new words, add an optional voice reference, and generate takes. Acting presets help steer the delivery, or you can import your own recording
  • Variations renders several seeds and ranks them by how many of the expected words local Whisper hears
  • Match level to nearby speech adjusts the new line by up to 6 dB so it blends in

Fix the lips

  • Pick the speaker on the first frame, track their face with FaceRefine, then draw an oval over the mouth
  • Tune denoise strength, mouth mask softness and face color match, then compare the original and the repair side by side
  • Each lip repair is limited to about seven seconds

Three repair modes

  • Voice only: fix the speech and keep the original picture
  • Off-screen voice: the same, for a speaker who isn't in frame
  • Joint retake: H3 regenerates the voice and the mouth together in one pass, with the real picture and sound around the interval kept as context. It needs a window of at least 124 frames, so your source needs about 5.2 seconds of video. Audio edges land within about 25 ms and video edges within about 170 ms

For dialogue scenes you can add up to eight turns per clip, reuse a voice by giving turns the same speaker label, render them all in order, and export everything together. Each finished lip repair can also download its ComfyUI API-format workflow.

Setup

Start with an up-to-date ComfyUI that already runs MiniMax H3, then clone the studio into custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/Nebuluss/H3-AudioRetake-studio.git

You'll also need:

  • The ComfyUI-H3-Motion-Context-MultiRef, ComfyUI-H3-FaceRefine and ComfyUI-VideoHelperSuite node packs
  • FFmpeg and FFprobe on ComfyUI's PATH, plus OpenCV and NumPy for face tracking
  • H3 Ref2VA weights, the Qwen H3 text encoder, the H3 video and audio VAEs, and a Pika H3 Flex Ref2VA LoRA in the right model folders. Pick the filenames in the studio's H3 model setup panel. Nothing downloads automatically
  • A Hugging Face-format Whisper model under models/whisper if you want Transcribe and Variations ranking

Restart ComfyUI and open http://127.0.0.1:8188/audio-retake/. The repo bundles an experimental six-step Pika sampler node that follows Pika H3 Flex's fast schedule.

Limits to know

  • It's early experimental software, tested only on a local Windows and NVIDIA setup so far
  • The model can change a face or add artifacts, even with a face reference. Check every result
  • Replacing an audio interval replaces everything you can hear there, including other voices and ambience
  • Exports are re-encoded, so untouched pixels and audio aren't bit-identical. Your original files are never overwritten
  • Speaker labels are manual. There's no automatic voice separation

License check

The studio is MIT. MiniMax H3 uses the MiniMax H3 Community License, whose grant excludes the EU, the UK, South Korea and the United States, and Pika H3 Flex has its own model license. Read each card before you use the results commercially.

Conclusion

Dialogue fixes have been one of the most annoying parts of AI video: one wrong word and the whole clip goes back into the oven. H3 Audio Retake Studio turns that into a quick, local edit, and it's a lovely sign of how fast the open H3 toolbox is growing.

Want to make an H3 clip with sound first? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=h3-audio-retake-studio-comfyui-dialogue-lip-sync-repair

Original Source

Go give your characters a better line!

—Aurelia ♡