MiniMax H3 R38 Refine Sampling

The second pass of IAMCCS's latent upscale isn't a full regeneration - it's a refine. The upscaled latent is already close to done; this node builds the sampling machinery that gently improves it at the higher resolution. It doesn't sample anything itself. Like the classic ComfyUI SamplerCustomAdvanced pattern, it outputs a noise, a sampler, sigmas, and a patched model that you feed into an actual custom-sampler node. The "prep the pieces" separation lets the same refine pipeline be driven by any downstream sampler.

How it works

• Derives a per-segment seed: plan seed + segment_index × seed_stride + a fixed 10000 offset, masked to 64 bits. Same seed math every run, so a given segment of a given render is reproducible.
• Reads refine steps and denoise from the plan's upscale settings - low steps and a low denoise (around 0.2) are the whole point. High denoise would destroy the base and you'd be regenerating, not refining.
• Applies the plan's turbo LoRA first, then any secondary LoRA, then MiniMaxH3SigmaShift with the video/audio shift from the sampling config (12 / 3 defaults).
• Patches attention acceleration - ComfyKitchen's per-model attention when the plan says so, otherwise the plan's native acceleration path.
• Builds RandomNoise, KSamplerSelect (default sa_solver) and BasicScheduler (default simple) - all ComfyUI-core nodes, just called from here so the settings are centralized.

The report string is unusually informative: it logs the effective seed, steps, sampler/scheduler, denoise, turbo status, secondary LoRA status, and which acceleration ran. If you're chasing a variable-quality repro, read that log line first.

Inputs

• model - the H3 model (turbo/secondary LoRA get applied inside).
• cine_linx - the shot plan with upscale settings.
• segment_index - forces the per-segment seed math.

That's it. Everything else (sampler name, steps, denoise, shift, seeds) is read from the plan, not exposed as widgets here.

Outputs

• model - patched and attention-accelerated, feed into the sampler node.
• noise, sampler, sigmas - the three sampling components for SamplerCustomAdvanced.
• report - the status line.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Requires MiniMax H3 support in ComfyUI, the plan's turbo/secondary LoRAs in your loras folder if configured, and (only if the plan requests it) the upscaler checkpoint and ComfyKitchen attention path.

Gotchas

• If the seed doesn't seem to do anything, remember it's offset per segment by segment_index × stride + 10000. The visible seed in your settings is only the base.
• This is refine machinery for the latent upres route. Wire it into a workflow whose plan says native-only and it'll happily build sigmas the downstream never asked for - the report will tell you if upscale is actually on.
• Missing turbo LoRA that the plan expects → the turbo application path can error or silently skip; keep the LoRA folder in sync with the plan.