MiniMax H3 Image • Sampling Preset

H3's sampling recipes are the kind of thing you'd spend a week getting wrong on your own. Base quality is res_multistep with a simple scheduler at 20 steps and a video/audio sigma shift of 12/3. Turbo is euler/simple at 8. These aren't values you can intuit, and getting the shift pair wrong produces mush. H3ImageSamplingPreset - "MiniMax H3 Image • Sampling Preset" - packages the documented recipes so you pick a name and get the whole correct configuration: sampler, scheduler, step count, and the H3 sigma shifts, all applied to your model.

What it returns

Feed it your loaded H3 diffusion model in model, pick a sampling_profile, and it returns everything your sampler needs: a cloned model patched with the recipe's video/audio shifts, a ready-made sampler, a complete sigmas schedule, and a sampling_info string recording what got resolved. The profiles are the ones from the README's sampling table - base quality | RES 20 steps (the quality reference), base speed | RES 12, the official Turbo v1.0 | 8 steps, Turbo v1.0 768p | 4 steps, REF2VA Turbo v0.1 | 4 steps, and the experimental hybrid single image | ER-SDE 8 steps that reproduces the community's one-frame workflow.

The honest advice on which to pick: base quality at 20 steps is the reference you compare everything against, and the official Turbo v1.0 eight-step profile is the practical speed/quality default. The 768p four-step profile is when you're chasing speed specifically. Start there, not at maximum quality | 20 frames in the prepare nodes - those frames buy temporal context, not raw fidelity.

The Turbo adapter rule

Two things the README is emphatic about. First, the Turbo profiles require the exact matching LoRA adapter, and this node does not load it - you load the adapter separately into the pipeline from ComfyUI/models/loras/. FL2VA Turbo v1.0 wants minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors, the 768p profile wants the 4-step 768p adapter, REF2VA Turbo wants the REF2VA 4-step adapter. Don't mix FL2VA and REF2VA adapters, and don't reuse one profile's shifts with another adapter - both are explicitly warned against.

Second, the custom | use controls below profile. It exposes custom_sampler (44 samplers), custom_scheduler, custom_steps, custom_denoise, custom_shift_video, custom_shift_audio, and beta-scheduler parameters. But those controls are ignored while a documented preset is selected - deliberately, so that loading an existing workflow always preserves its exact recipe. You only get the custom controls when the profile literally says custom.

The stale-backend gotcha

One gotcha that will bite if you've been running an older version: if a profile like Turbo v1.0 | 8 steps is missing from the dropdown, your node definitions are stale - the files on disk updated but ComfyUI is still running the old classes in memory. Update the pack, stop every ComfyUI process, start it again, and reload the browser. Don't try to fix it by hand-typing the rejected values into a sampler; the workflow wants the new backend.

Install

Install is the pack: ComfyUI Manager (search MiniMax H3 Image Studio) or git clone https://github.com/astropuzzo/ComfyUI-MiniMax-H3-Image-Studio.git into custom_nodes, then restart. No extra Python dependencies - it builds on ComfyUI 0.30.0+'s native H3 sampling - but you do need the H3 diffusion model and any Turbo adapter you plan to use.