H3 Retake (redo part of a clip)

H3 Retake redoes one stretch of a finished H3 clip and keeps everything else. You load the clip's frames and audio, set a time window, write a prompt for that moment, and only the material inside the window is regenerated - the rest is frozen as raw latents. Picture and sound are independent, so you can redo the video while keeping the original performance, or keep the picture and redo the audio.

This is the node you want when a 30-second chain comes back with one botched moment - a face that broke, a line that garbled - and the thought of rerendering the whole take makes your stomach drop. Rerendering the chain rerolls every join. A retake reopens only the bad window.

How it works

It encodes both sides of your finished clip to raw latents - video through the video VAE, audio through the audio VAE - and builds a latent with a noise mask: zeros everywhere except the window you're redoing. Then it samples that window with a BasicGuider at cfg 1.0 (H3's normal setup, no negative branch), decodes, and re-appends any off-grid tail untouched.

The inputs are the full rig: model, clip, video_vae, audio_vae, the finished clip's images and audio, a prompt for what should happen in the window, start_seconds/end_seconds, mode, seed, and - importantly - the same sampler and sigmas the original render used. Wire in the exact schedule you rendered with (euler / beta, the sigma-shift chain from the render canvas) or the retake won't blend.

mode is the interesting control, three ways:

• video + audio (redo the moment) - the default, both halves.
• video only (keep the performance) - new picture, the original line stays.
• audio only (keep the picture) - redo the sound without moving the image.

The output is images, audio, and an info string that reports what was redone, which slots were frozen, and how long it took.

Installing it

Part of the ComfyUI-H3-Multishot pack:

cd ComfyUI/custom_nodes
git clone https://github.com/jlucasmcrell/ComfyUI-H3-Multishot

Or search H3 Multishot in ComfyUI Manager. Requires ComfyUI v0.30.0+, the H3 checkpoint, and both the video and audio VAEs from Comfy-Org/MiniMax-H3.

Gotchas

Three things to respect. First, start_seconds snaps to H3's latent grid (roughly 0.14 s per slot) - the node logs the actual window, so check it rather than trusting the seconds you typed. Second, frozen material isn't a prompt; the model sees only your retake prompt plus the raw latents either side, so write the prompt as if it's the shot at that moment, not a description of the whole clip. Third, if the retake doesn't visually match the surrounding footage, the usual culprit is a schedule mismatch - sigmas and sampler must match the render canvas or the window's noise character will differ from the frozen material and the seam will show.