MiniMax H3 Guide Pass

It is not a refine, and that distinction is the whole node

Everyone's mental model for "make my clip better" is the second pass: re-noise the latent you already have partway down the schedule and let the sampler resolve it. That's H3's own two-pass upscale, ReDetail on LTX 2.5, img2img at low denoise.

"MiniMax H3 Guide Pass" does something else. It reads every generated shot on the reel back off disk, encodes each one, pins it into the model's own timeline as an aligned guide, and generates the shot again from noise - the entire schedule, full denoise. Nothing of the old latent survives. What comes back matches the old video because the file it runs under was trained to match it, pixel for pixel - not because the sampler was forbidden from wandering.

That's the point of a guide LoRA: not a description of a look, a correspondence. Guide token (t, y, x) sits on target token (t, y, x), so you get a real sharpener or style transfer instead of a re-roll that drifted. Think ControlNet's idea, with your own clip as the condition image.

How it works, in order

For each generated part on the reel: frames are read back, encoded by the video VAE, and pinned as one guide block at frame 0 of an empty AV latent at the same canvas and length - core's MiniMaxH3AddGuide with resolved_frame_index: 0, exactly what the published guide files were trained against. A seam-trimmed shot is padded up to H3's 17k+5 frame grid with its own last frame and trimmed back after decoding. Then the sampler runs the full schedule from noise, the picture half is decoded, and the audio half is thrown away - the new shot is rewritten against the soundtrack the old one already had, so the sound you judged the picture by is the sound you keep.

The guide files are Ref2VA adapters, so the pass runs on Ref2VA whatever the rest of the piece routed to, wearing the published rig: the turbo distill the files were trained against plus the guide file, on its own sampler row. The piece's own LoRAs stay off - they drew the shot; this pass maps it.

Inputs and outputs you actually touch

model is the checkpoint already wearing the pass's stack; clip and vae are the H3 text encoder and video VAE; reel is the pack's own reel type, which is why you can't hand-wire this node into an unrelated graph. prompt is what the file is told: the sharpener's published caption, or the style itself for a style file.

steps defaults to 8 and cfg to 1.0 - H3's checkpoints are CFG-distilled, so guidance on top burns the picture and doubles the cost - plus sampler_name, scheduler and seed. Seeds are per shot: each part's position is added to yours, so one number covers the reel without every shot landing on an identical sample.

picture is optional and only the style file reads it: a look's frame as atlas:000123, or any file under input/, presented to the tokenizer as <Picture 1> beside the guide. The output is a single reel, each generated part replaced by its re-rendered version.

In a graph it sits after the last shot is written, before ReDetail and the neural refiner, never inline at a seam: a sharpened tail handed to the next shot as its anchor compounds - the ratchet the DLSS refiner was measured to have. The feathered-seam join is the author's own open measurement.

Install

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

Restart. No pip install. Then the files: the two guide LoRAs (minimax_h3_lms_v1.0_r64.safetensors, minimax_h3_style_transfer_v1.0_r64.safetensors) from Alissonerdx/Minimax-H3-ComfyUI and a minimax_h3_ref2v_turbo_4step distill, all into models/loras, plus the H3 checkpoints, text encoder and VAEs. Nothing is downloaded for you.

Common issues

• "A reel with nothing generated on it - there is nothing to re-render." You pointed the pass at a reel with no generated parts. A reel of held takes passes through untouched instead - assembling clips on a machine that only wants to save them shouldn't need the weights.
• The style barely comes through. Check that the published ref2v…turbo distill is installed. With the piece's own turbo file standing in, the pass is measurably under-driven - style by half, sharpen hardly at all. That distill is the pass's rig, not decoration.
• Wrong VAE, loud failure. Wire the H3 video VAE, not the audio one; the pass raises on a shape mismatch rather than quietly producing mush.
• Raylight. Two-GPU rendering loads the checkpoint inside Ray's workers, which this pass can't sample across. Switch it off or render on one card.
• Cost and canvas. One extra full generation per shot, on guide files whose examples were made around 0.59 MP. Past that, the author calls it unmeasured - test on your own footage.