<a href="https://www.youtube.com/watch?v=A53fIhsyTm8"> <img src="https://img.youtube.com/vi/A53fIhsyTm8/maxresdefault.jpg" alt="MiniMax-H3 ComfyUI Tutorial" width="750"> </a> ▶️ Click the image above to watch the complete step-by-step walkthrough on YouTube.

Features

One main node (H3 Image (Simple)): prompt, up to 9 reference images, size, sampling and decode in one place.
Exact resolution: pick an aspect ratio and megapixels, or type any custom width and height. With exact_size on, the output is exactly the size you typed.
No hidden passes: no automatic refine pass and no automatic upscale.
Multi-reference support: connect images to ref_image_1 to ref_image_9 and use <Picture N> in the prompt.
Optional detail pass: an upscale model adds fine detail (faces, textures), then the image is shrunk back to its original size. The resolution does not change, and detail_strength sets how much is blended in.
Faster iteration: prompt and reference encoding is reused when only the seed or sampling settings change.
Final Size node (optional): exact resize, scale, crop or pad, with an optional upscale model.

Requirements

A ComfyUI build with the native MiniMax H3 nodes (comfy_extras/nodes_minimax_h3.py)
Your H3 model, text encoder and VAE. A turbo LoRA and an upscale model are optional.

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/SatoDive/ComfyUI-H3-IMG-Gen-SatoDive.git

Or download the ZIP and extract it into ComfyUI/custom_nodes/. Restart ComfyUI afterwards. __init__.py must sit directly inside the ComfyUI-H3-IMG-Gen-SatoDive folder.

Nodes

H3 Image (Simple) - SatoDive

| Input | What it does | | --- | --- | | model, clip, vae | Your H3 model, text encoder and VAE | | prompt | Your prompt. Use <Picture N> to point at references | | size_mode | Aspect + megapixels or Custom size | | aspect, megapixels | Used in Aspect + megapixels mode | | width, height | Used in Custom size mode | | exact_size | The model works in multiples of 32. When on, the result is resized by the few leftover pixels to match your typed size exactly | | seed, steps | Sampling. Match steps to your LoRA: a 4-step LoRA needs about 4, other turbo LoRAs about 20, no LoRA about 50 | | lora_name, lora_strength | Optional turbo LoRA | | sampler_name, scheduler | Sampler settings | | detail_model, detail_strength | Optional detail pass. The resolution stays the same | | ref_image_1 to ref_image_9 | Optional reference images | | ref_image_size | max = best likeness, slower and heavier on VRAM. match = faster, weaker likeness |

Outputs: image, width, height (the real size of the image).

H3 Final Size - SatoDive

Optional last step:

Exact size: stretch, crop or pad
Scale by factor
Target megapixels: keeps the aspect ratio
Keep size (detail only): adds detail with the upscale model, then shrinks back to the current resolution

Workflows free to download :

https://www.patreon.com/SatoDive/posts/minimax-h3-is-171110570

Tips

Faster iteration: 4 steps with a 4-step LoRA, ref_image_size on match, no detail model, lower resolution. Switch to the heavier settings once the composition is right.
Detail pass: an upscale model sharpens and adds texture, but it does not redraw faces. At very low detail_strength it does almost nothing.
VRAM: a 4x upscale model turns a 3 MP image into a roughly 48 MP intermediate, which is slow and memory-hungry on small GPUs.

No other custom node pack is needed. The single-frame still latent and still decode are built in (adapted from ComfyUI-Fizgig-H3-Still, MIT).

Status

Work in progress. Report problems in the Issues tab.