MiniMax H3 Image • Advanced Sampling

When H3ImageSamplingPreset isn't enough control, this is where you end up. H3SamplingSettings - "MiniMax H3 Image • Advanced Sampling" - exposes every dial directly: sampler, scheduler, steps, denoise, and the two sigma shifts that are the actual secret to making H3 behave. It's the node for people who know what they're doing, or for the rest of us who are experimenting one parameter at a time.

Why the shifts matter

The mechanism is the interesting part, because it's not just "set some numbers." H3 is a FLOW_AV model in ComfyUI - the packed latent contains an audio stream even during image-only denoising - so the node clones your model and patches its sampling object to keep ModelSamplingAV semantics alive, then sets shift_video and shift_audio as flow sigma shifts. On current ComfyUI it uses the native AV sampling; on older H3-capable cores it falls back to a compatibility shim. Either way, sampling_info tells you which backend you got.

The dials

Inputs, and which ones you'll actually touch:

• model - your loaded H3 diffusion model.
• sampler_name - defaults to res_multistep, which is the H3 baseline. That's the honest default; treat deviations as experiments.
• scheduler - defaults to simple; beta_custom is added on top of ComfyUI's list and exposes beta_alpha/beta_beta (both 0.6).
• steps - the number actually executed (20 is the quality reference).
• denoise - ComfyUI-style strength. Below 1 it builds a longer schedule and keeps only its final steps; 0 returns empty sigmas (useful for pure latent passes).
• shift_video (12) and shift_audio (3) - the flow sigma shifts for the video/image and audio streams. This is the pair that breaks workflows when it's wrong.

Outputs: shifted_model (the patched clone), sampler (for SamplerCustomAdvanced), sigmas (the full schedule after steps/scheduler/denoise are applied), and sampling_info.

The adapter trap

Here's the trap the pack keeps warning about: the shift values and sampler must match the adapter you're using. The official Turbo recipes are euler/simple at 8 steps with shifts of 12/3; base quality is res_multistep/simple at 20 with 12/3. Reusing a Turbo adapter with a base profile's settings, or mixing FL2VA and REF2VA adapters and their shifts, produces garbage that looks like a model problem when it's a settings problem. The README's advice applies to every knob here: change one optimization at a time and compare with the same seed. Stacking caches, attention patches, and distilled adapters together is how people end up with a graph where nothing works and no single change fixes it.

A note on shift_audio: it's metadata the model consumes during sampling, but Image Studio never decodes the audio VAE - you're generating a still, the audio stream is just along for the ride in the latent. So don't chase the audio shift for image work; leave it at 3 unless you're deliberately porting a video recipe.

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. Requires ComfyUI 0.30.0+ (the native AV sampling path) and the H3 diffusion model; no extra Python dependencies.