IAMCCS MiniMax H3 Progressive Spatial Sampler

This one is a sampler replacement, and if that's not what you were expecting from an IAMCCS node, fair enough. Look at its input sockets - noise, guider, sigmas, latent_image - and you'll recognize the exact contract of ComfyUI's SamplerCustomAdvanced. You can drop it into an advanced-sampler slot and it will do the same job, except it won't denoise at full resolution the whole way through.

The trick is progressive spatial sampling, aka a coarse-to-fine ladder. Instead of one full-res pass, it solves the broad spatial structure on a downscaled video latent first, then bumps the video resolution up as the sigma schedule descends toward the detail-forming steps. Two clever details make it work. First, the audio latent is carried through every spatial transition untouched - byte-for-byte, per the code - which is the right call on H3, where video and audio live in one nested latent and audio is the expensive thing to keep coherent. Second, each resolution bump runs a "high-frequency re-entry" step that injects fresh high-frequency noise, scaled by the current sigma, so the upsampled latent doesn't just look like a smeared blow-up of the low-res result.

It deliberately uses stock ComfyUI guider/sampler plumbing (an Euler core) plus PyTorch interpolation. The author is explicit that it's an independent implementation, not a copy of anyone's third-party progressive sampler.

What you set

Really just one thing: profile.

• iamccs_progressive_2stage (default) - a 0.5→1.0 ladder. The author positions this as the daily-production baseline because it stays closest to a normal full-resolution pass.
• iamccs_progressive_3stage - a ⅓→⅔→1.0 ladder. Saves more early spatial work, but the code itself warns it can change fine text, small props or prompt adherence. Treat it as an experiment.
• iamccs_progressive_pdd_2stage - the same two-stage ladder while the native PDD head bank owns the eight-step denoise schedule.

Outputs are output (LATENT), denoised_output (LATENT) and report (STRING). Worth knowing: this class returns the finished latent on both output slots, and the report strings together a stage breakdown like 512x288:4step -> 1024x576:4step, which is genuinely handy for confirming the ladder did what you expected.

When you'd actually reach for it

This lives under IAMCCS/Shotboard/Backends/MiniMax H3, which tells you where it belongs: as a swappable backend inside the IAMCCS Shotboard pipeline, not as a general-purpose sampler for every H3 graph. If you're doing long or high-res shots and the full-res early steps are the expensive part, the 2-stage profile is a reasonable speed experiment that keeps a close relationship to your normal pass. It also needs a real MiniMax H3 nested AV latent (video + audio) - feed it a plain latent and it raises immediately.

Two honest caveats. It does not support noise_mask yet - the code rejects masked sampling and points you to the native/PDD path. And despite the "progressive" branding, don't expect a miracle quality win; this is a compute-saving optimization, not an upscaler. Where people get burned is reaching for the 3-stage profile because it sounds faster and then chasing prompt-adherence regressions for an afternoon. If you're experimenting, 2-stage first, always.

Install

Part of IAMCCS-nodes, so the usual:

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

or install via ComfyUI Manager (search "IAMCCS"), then restart. You'll also need the MiniMax H3 model and VAEs loaded through ComfyUI's native H3 support - this node orchestrates H3 latents, it doesn't ship them. And remember the H3 licence geography: the community licence excludes the US, EU, UK and Korea. Current ComfyUI (≥ 3.x) with Python ≥ 3.12 and PyTorch ≥ 2.8 is the stated floor.