Configuring MiniMax H3 Image in ComfyUI Against Artifacts
MiniMax H3 Image is a distilled diffusion transformer designed for generating ultra-high-resolution images up to 16 megapixels. The model uses a six-step schedule and does not require a negative prompt.
When run by default at low resolutions or with standard samplers, an over-sharpened texture effect often occurs. Granular and blocky artifacts, known as the 'crunchy' grid, appear on skin and fine details.
The problem is solved by three adjustments in the ComfyUI pipeline without needing to replace the model itself.
To eliminate defects, perform the following settings:
-
Increase native resolution closer to the target (8–16 MP+) followed by downscaling instead of compressing details into a low resolution.
-
Switch to the SamplerLCM node with the Simple scheduler at 6 steps and reduce the s_noise parameter from 1.10 to 0.90 to smooth micro-noise.
-
Abandon tag sets in favor of coherent descriptions and connect community patches, such as the MiniMaxH3PromptWeight node and a single-frame decoder instead of the base video VAE.
The model runs comfortably on GPUs with 16 to 24 GB of VRAM. For devices with less memory, quantized GGUF versions with offloading can be used.










