MiniMax H3 Drift Guard

Shoot an H3 video where every shot continues from the last one and, by shot five or six, something starts to look wrong. Colors creep brighter, fine detail gets that overcooked, over-sharpened crunch. It's not one bad shot - it's drift that compounds, because each continued shot inherits the previous one's flaws and adds a little more. That's exactly what this node exists to stop, and it does it without retraining anything.

The H3 Drift Guard is ComfyUI-Continuity's fix for a real quirk of how H3 (and any rectified-flow sampler) hands you a finished frame. At every denoising step, the model makes a full guess at the clean picture - the denoised the sampler steps toward. On a run that ends on an Euler step to zero, the frame you actually get is basically the last of those guesses. And the late steps are where the trouble lives: they add texture, and with it the over-sharpening and the small brightness bias that each continued shot inherits and then re-adds. Averaging the model's last few guesses instead gives you a steadier picture that carries less of that bias across the seam. The idea comes from work on error-free long-video generation; this node is that idea applied to H3 as a model patch.

The one knob: guesses

• model in, MODEL out - a patched clone you drop between your checkpoint and the sampler.
• guesses (default 3, range 0–99): how many of the schedule's last steps to average. One is the plain render - the sampler's own output. Zero means average every step's prediction, which is where you end up if you ignore the "don't crank it" advice: the early steps are the model's generic picture, so too many of them loosen faces and objects. Three is the sweet spot.

Two deliberate design choices worth knowing. It counts in guesses, not sigma, because H3 samples at flow shift 12 - a sigma window that means "the last three steps" on a 20-step schedule catches almost nothing on a turbo schedule, while "the last N guesses" means the same thing on both. And it averages the model's own predictions rather than the paper's noise − velocity form, since those two only agree when every step ran on one consistent model - which a turbo run (a few lead-in steps on base weights, the rest distilled) is not.

Using it

If you live inside Continuity's one-node UI, you don't wire this at all - you flip the drift guard setting on the pack's settings page. It's off by default, deliberately, and the page offers up to 8 guesses. It only applies to a pass and its lead-in: the refine, face-crop and restore passes resume partway down the schedule and redraw texture on purpose, so they never see it. And only the picture is averaged - H3 packs its audio into the latent, and the sound row comes out exactly as the sampler made it.

If you build H3 graphs by hand rather than through the front door, this is one of the few internal nodes you can actually reach for: search "MiniMax H3 Drift Guard" and put it between your model and the sampler. It's H3-only - an LTX or Wan chain won't know what to do with it.

Installing it

Same story as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/roadmaus/ComfyUI-Continuity

Restart, nothing to pip install. You need a recent ComfyUI - H3 ships in core now - plus H3's weight files (MiniMax-H3 on HuggingFace). Two folder gotchas: this pack was ComfyUI-MiniMax-Creator before the rename, and cloning the new folder beside the old one registers duplicate ids so no node appears at all - git pull inside the old folder instead. And H3's community license is geofenced; if you're in the US, EU, UK or South Korea, running the local weights isn't licensed, and no node changes that.

Where people get burned: expecting the guard to fix a bad prompt, or dialing guesses to double digits and wondering why everyone's faces went soft. Keep it at 3–5, let it do one job, and the frying stays out of your next seam.