MiniMax H3 Image ComfyUI: T2I & edit workflow, model files, settings

ComfyUI ships two official templates for MiniMax-H3-Image: Text to Image and Image Edit. Both need ComfyUI 0.45.0+ and template pack 0.11.104+. Search "MiniMax-H3-Image" in the Template Library; if it is missing, update ComfyUI.

Model files

All four come from Comfy-Org mirror repositories on Hugging Face.

Role	File	Size	Folder
Diffusion model (pruned INT8 DiT)	`minimax_h3_image_pruned_int8_convrot.safetensors`	21 GB	`models/diffusion_models/`
Text encoder (Qwen3-VL-32B INT8)	`qwen3vl_32b_minimax_h3_int8_convrot.safetensors`	25.28 GB	`models/text_encoders/`
Prompt expander (Qwen3.5 9B W4A8)	`qwen3.5_9b_w4a8_gptq.safetensors`	7.34 GB	`models/text_encoders/`
VAE (FP16)	`minimax_h3_image_vae_fp16.safetensors`	1.17 GB	`models/vae/`

Total download: about 54.8 GB with the expander, 47.5 GB without it (sums of the sizes above).

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── minimax_h3_image_pruned_int8_convrot.safetensors
    ├── text_encoders/
    │   ├── qwen3vl_32b_minimax_h3_int8_convrot.safetensors
    │   └── qwen3.5_9b_w4a8_gptq.safetensors
    └── vae/
        └── minimax_h3_image_vae_fp16.safetensors

Sampler settings

Setting	Value
Sampler	euler
Scheduler	simple
Steps	6
Guidance (CFG)	1
Denoise	1

CFG is 1, so there is no negative prompt. Write what to avoid into the positive prompt.

VRAM and resolution

The Text to Image template opens at 4 MP on a square canvas, below native resolution on purpose, so low-VRAM machines do not run out of memory. Raise the megapixels with the resolution selector when your GPU has room. The ComfyUI guide does not publish a minimum VRAM number.

Aspect ratio	Native ~16 MP size
1:1	4096 × 4096
4:3 / 3:4	4736 × 3552 / 3552 × 4736
3:2 / 2:3	5024 × 3360 / 3360 × 5024
16:9 / 9:16	5472 × 3072 / 3072 × 5472
21:9	6272 × 2688

Edges are generated on multiples of 32 px. If an edit runs out of memory, enable the template's Resize Image/Mask node and lower its megapixels.

MiniMax H3 Image edit workflow (multiple references)

The first loaded image is the one being edited; extra images supply identity, props, materials or composition, up to 16 references. To add one, place another Load Image node and connect it to both the prompt generation step and the conditioning node. ComfyUI rescales the internal copies itself.

The prompt expander

Both templates run a text-generation step before the sampler. It rewrites a short request into the detailed English caption the model reads best and can pick the aspect ratio. In Text to Image it is off by default; in Image Edit it is on. The official expander is MiniMax-H3-Image-PE (Qwen3.8-27B, FP4); the templates use a lighter Qwen3.5 9B build so it fits beside the 21 GB DiT on one GPU.

More: run it without ComfyUI · prompt guide.

MiniMax H3 Image in ComfyUI