H3 Studio • Sampling Preset

Sampling H3 is not "pick Euler a, set steps to 20, go." H3 image work runs on specific recipes - the model was tuned for a video setting and the community's fast profiles are a fragile collection of adapters, empirical settings and special samplers. H3StudioSamplingPreset exists so you don't have to memorize which sampler and scheduler go with which step count.

Feed it your loaded H3 diffusion model and pick a sampling_profile. It hands back ready-to-use model, sampler, and sigmas outputs plus a sampling_info string - the "SAMPLER" and "SIGMAS" types are custom, so this node plugs into a custom sampler setup rather than a stock KSampler.

The profiles

The eight choices are clearly a ladder with the training wheels labeled:

• base quality | RES 20 steps (default) and base speed | RES 12 steps - the safe native profiles using RES Multistep. Start here.
• LightX v0.1 | ER-SDE 4 steps and LightX v0.1 | SA-Solver 4 steps - these reproduce Kijai's empirical four-step LightX v0.1 recipes, which are faster but older and, being empirical, less "official" than the later v1.0 adapter. The node's description is upfront about this, which I appreciate.
• quality | 20 steps and speed | 12 steps - generic non-RES fallbacks.
• turbo | 8 steps (LoRA) and turbo | 4 steps (LoRA, experimental) - the LightX LoRA-driven profiles.

Two honest gotchas live in the details. First, the LoRA itself must be loaded upstream - this node applies the sampling recipe, it does not load the LightX adapter for you. Wire a LoRA loader before it if you're using a turbo/LightX profile. Second, the v1.0 LightX path in the bigger pack picture uses the full 8-step adapter; the v0.1 profiles here are the empirical ancestors, so don't expect identical results to the Director's lightx_v1_fl2v_8.

When to reach for it

If you're building a custom H3 sampling graph instead of using the maintained workflow, this is the node that gives you the "proper" H3 settings without reverse-engineering them. It's the explicit, no-magic version of the sampling logic the Director and Context Sampling Preset wrap. The sampling_info string is worth reading the first time - it tells you exactly which sampler/scheduler/sigma recipe got applied, which is how you'll know you're on RES Multistep and not accidentally on something experimental.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/thaakeno/ComfyUI-MiniMax-H3-Studio.git
cd ComfyUI-MiniMax-H3-Studio
python -m pip install -r requirements.txt

Restart and hard-refresh the frontend. No extra pip dependencies. Alpha software, so treat the turbo/LightX entries as "works, mostly" and keep base quality in your back pocket for when a fast profile does something weird.