Simple t2i & i2i workflow for Minimax H3 Image
A streamlined text-to-image workflow for MiniMax H3 Image, featuring live preview via tiny VAE so you can see the generation progress in real time.
Version Notes
v1.0 = simple text to image (t2i) workflow with preview feature
v1.0plus = v1.0 + prompt enhancement + preset sampler combo
v1.0pro = v1.0plus + Image Reference & Editing (i2i)
v1.0max = v1.0pro + 16 refs supported with performance optimization
Key Features
Real-time preview — Powered by
ModelPreviewOverrideKJwithtaeh3_imagetiny VAE (4.96 MB), no waiting until the end to see resultsPrompt enhancement — available in v1.0plus/pro and can be disabled manually.
Preset Sampler Combo — available in v1.0plus/pro. for easy switching.
Image Reference & Editing — available in v1.0pro/max. for greater creative control (3 reference images for v1.0pro, 16 refs for v1.0max with performance optimization)
Tips
The preview node requires the tiny VAE file — don't skip it!
Default resolution is set to 4096×4096 (16 MP, 1:1 aspect). Adjust via the ResolutionSelector node. (For v1.0pro, please pay attention to the toggle switch between custom resolution and the reference image's resolution)
euler/simple/6steps: Officially recommended settings.
sa_solver/sgm_uniform/5steps: If you find the image too noisy or notice excessive smoothing/smearing (especially when rendering realistic human skin), you might want to try this sampling configuration, which I personally recommend.
Required Models
minimax_h3_image_pruned_int8_convrot.safetensors
qwen3vl_32b_minimax_h3_int8_convrot.safetensors
minimax_h3_image_vae_fp16.safetensors
taeh3_image.safetensors
qwen3.5_9b_w4a8_gptq.safetensors (for prompt enhancement only)
Important note! A friendly reminder:
You can use the text encoder from Moss Image as a replacement for qwen3.5_9b_w4a8_gptq.safetensors in the role of the PE model.
for example, the one I recently uploaded: Qwen3.5-9B-Heretic2.1
Required Custom Nodes
ComfyUI-KJNodes — for
ModelPreviewOverrideKJ……rgthree-comfy — for
Power Lora Loader,Seed……
Installation
Install required custom nodes via ComfyUI Manager → Custom Nodes Manager
Download all models into their respective
ComfyUI/models/subfolders (see Model Links node in the workflow for paths)Download
taeh3_image.safetensorsintomodels/vae_approx/Load the workflow and queue prompt
Description
Plus what?
prompt enhancement
preset sampler combo for easy switching


