MiniMax-H3 Turbo

If you've loaded a MiniMax-H3 workflow and seen a sampler node with zero inputs and zero idea what to do with it, this is that node. MiniMaxH3Turbo is the step-count drop you need for H3's Turbo LoRA: a custom SAMPLER that collapses a generation that could take dozens of denoising steps down to four. H3 is MiniMax's multimodal video model - it emits video and native stereo audio from one latent - and like every distilled "turbo" line before it, the fast regime only works with the sampler it was trained against.

What it is

It's a sampler, not a workflow. You feed it into a SamplerCustomAdvanced the same way you'd wire a regular KSampler's sampler port, and it does the denoising for you. The author's own description is the whole manual: "4-step sampler for the MiniMax-H3 Turbo LoRA. Feed into SamplerCustomAdvanced and set the scheduler to 4 steps." No inputs, one SAMPLER output, nothing to configure.

How it works

This is the interesting part. H3 doesn't produce a plain video latent - its latent packs the video channels and the audio channels into one tensor, side by side. A stock sampler steps that whole thing on one sigma schedule and the audio comes out mangled. MiniMaxH3Turbo splits the tensor on the fly, steps the video portion on the regular schedule, and steps the audio portion on a separate, slope-scaled schedule (_audio_sigma / _audio_slope in the source). It also expects to be given an H3 video+audio latent - the kind EmptyMiniMaxH3LatentAV or MiniMaxH3ImageToVideo produces - and raises a clear RuntimeError if you hand it a plain latent instead.

That split is the whole reason this node exists rather than you just picking "euler" and moving on. Distilled samplers are finicky; the audio half makes it worse. Letting the node handle both halves means one less thing to get wrong.

The settings that actually matter

• Steps: 4. The brief and the node description are explicit - set the scheduler/denoise to 4 steps. This is a Turbo LoRA, so like SDXL-Lightning or any other few-step distillation, you match the step count the LoRA was trained for. Running it at 20 steps doesn't give you more quality, it gives you a worse result than the distilled default.
• The Turbo LoRA must be loaded on the model, or "4 steps" is just a bad 4-step pass.
• The latent input to the sampler chain must be an H3 video+audio latent, not a regular LATENT.

Install

It ships in the PainterNodes pack, so it's one install for the whole set. In ComfyUI Manager, search "PainterNodes" and hit install. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/princepainter/ComfyUI-PainterNodes
cd ComfyUI-PainterNodes
pip install -r requirements.txt

Then restart ComfyUI. The pack's own requirements are feather-light (soundfile and numpy) - the heavy lifting, like the H3 model weights and the Turbo LoRA themselves, comes from wherever you loaded your H3 checkpoint, not from this repo. The author ships example graphs in the pack's workflows/ folder; when in doubt, load one and trace what connects to what.

Common issues

• "MiniMaxH3Turbo expects the MiniMax-H3 video+audio latent" - you wired a standard latent into the sampler. Check that the conditioning/latent builder upstream is the H3 one (the ref-to-video or FL2VA nodes from this same pack).
• Crunchy or garbled audio - almost always a steps/scheduler mismatch. Confirm 4 steps and that the Turbo LoRA is actually applied.
• Node won't load at all - the newer nodes in this pack target recent ComfyUI core; update ComfyUI before blaming the node.

The name sounds like marketing, but the mechanism is real: one tensor, two schedules, four steps. If you're on H3's Turbo LoRA, this is the sampler you want.