Difforum · Look Mix

MiniMax H3 and LTX render motion that actually makes sense. That is the problem. They also smooth away the grain, the boil and the per-frame shimmer that made Deforum, AnimateDiff and Disco Diffusion feel like something alive rather than something rendered. Look Mix is the node that puts the texture back.

The trick is that you render the same shot twice: once with a video model for the motion, once with the Feedback Sampler (or Restyle) for the look. Then you keep the video model's movement and light and donate the look pass's high-frequency detail to it. The look pass does not have to be the same length as the video - it gets resampled in time - so a cheap 12-frame look pass can texture a 120-frame H3 clip.

How it actually blends

Four modes, and they are genuinely different operations rather than intensity settings.

detail transfer is the useful one. The node blurs the look pass by detail_radius and subtracts the blur from the original, leaving only fine texture, then adds amount of that back onto the video frame. It's an unsharp mask pointed at the look pass instead of at the video. colour + detail does the same thing but first pulls each frame's palette toward the look pass in LAB, which is what you want when the two renders don't match tonally. flicker cuts does something cruder and more interesting: every flicker_every frames it swaps that frame harder toward the look pass, with seeded jitter - roughly one cut in six of the candidate frames is skipped, so the flicker doesn't land on a metronome. crossfade is a plain linear mix, and none passes the video through untouched.

step_fps is separate from all of it and works with no look pass wired at all. Set it to 8 or 12 and the node holds frames to that rate before output, which is the stop-motion / AnimateDiff stutter. It reads fps to know how much to hold, so set fps to the clip's real frame rate or the stutter comes out at the wrong tempo.

The inputs that matter

You wire the H3 or LTX decode into video and the Feedback Sampler's frames into the optional look_pass. Everything else is taste: start with blend on detail transfer, amount around 0.5-0.7, detail_radius 3, and flicker_every at something like 6 if you go that route. One output, frames, which goes to Upscale (2K / 4K) or straight to your video save node.

amount runs to 2.0 and that is not a typo - above 1 you're subtracting texture harder than the blur left behind, which pushes toward a crusty, posterised look. It's a legitimate effect on a music video and a mess on a face. seed only does anything in flicker cuts.

Install

ComfyUI Manager → search Difforum, or:

cd ComfyUI/custom_nodes
git clone https://github.com/chillithebillis/Difforum.git difforum

Restart, then find the node under Difforum/7 · Post. The pack's requirements.txt is numpy and nothing else, which ships with ComfyUI already. OpenCV is optional and only matters for optical flow and webcam features - Look Mix itself never touches it, so it works on a bare install. The pack is new (the author's first public release was mid-2026, and it barely has a Reddit footprint), so treat the info strings and the run reports as your documentation rather than searching for a tutorial.

Where people get burned

Detail transfer is a subtractive filter, so if the look pass is a different resolution or has compression mush in it, all you transfer is noise. If the result looks grey and grainy rather than textured, switch to colour + detail and lower amount before you blame the render.

The other trap is ordering. Look Mix sits after Restyle and before Upscale in the templates for a reason: upscale first and you're transferring texture into pixels that were interpolated, which bakes the shimmer in twice and gives you a file that looks crunchy at 2K and fine nowhere. And if your clip needs to be shorter than the render, cut before this node - a held frame inside the flicker pattern reads as a dropped frame, not as style.