MiniMax H3 Image: ComfyUI workflows for 8, 12 and 16 GB VRAM

Three versions of one ComfyUI workflow for MiniMax H3 Image. All three make full 16 MP images (3:4 = 3552x4736, 16:9 = 5472x3072), even on an 8 GB card. ComfyUI streams the model through the card in parts, so a smaller card gives the same image, only slower. Pick the file that matches your VRAM and drag it into ComfyUI.

File For Final image Draft VAE decode Reference pictures
minimax_h3_image_8gb_vram.json 8 GB cards 16 MP 4 MP Tiled (512) 10 slots, 4 or fewer is fastest
minimax_h3_image_12gb_vram.json 12 GB cards 16 MP 4 MP Tiled (768) 10 slots, 4 or fewer is fastest
minimax_h3_image_16gb_vram.json 16 GB+ cards 16 MP 4 MP Normal 10 slots

What you need for an 8 GB card

  • GPU: NVIDIA with 8 GB VRAM (tested: ComfyUI used at most 3.7 GB VRAM for a 16 MP image).
  • System RAM: 24 GB or more (tested with 24 GB), 32 GB is safer. The models are bigger than the card, so ComfyUI keeps them in RAM and streams them in parts.
  • Disk: about 38 GB free, on an SSD.
  • ComfyUI: a recent version (MiniMax H3 nodes are built in, no custom nodes).
  • 3 model files, the same full-quality files as the 16 GB version, no GGUF: the image model (21 GB), the main text encoder nvfp4 (15.7 GB) and the VAE (1.2 GB). Links are in Model files.
  • Not needed: the prompt upsampler qwen3.5_9b_w4a8_gptq.safetensors. It is switched off in the 8 GB workflow because it makes images very slow on 8 GB.

Measured render times

Tested on an RTX 5070 Ti (16 GB) with 24 GB of system RAM. For 12 GB and 8 GB, part of the card's memory was filled by a dummy program, so ComfyUI only had the VRAM a real 12 GB or 8 GB card has free. With the same seed, all three gave pixel-for-pixel the same image.

Job 16 GB 12 GB 8 GB
Text-to-image, 16 MP 66 s 63 s 60 s
Time per step (6 steps) 4.4 s 4.8 s 5.7 s
4 reference pictures, 16 MP - - 228 s (about 4 min)
4 reference pictures, 8 MP - - 156 s (about 2.5 min)

Real 8 GB and 12 GB cards (RTX 4060, 3060, 4070 ...) have slower chips than the test card, so expect longer times. The memory limit itself is not a problem.

Example images

Made with this workflow, straight out of ComfyUI with no editing or upscaling. Click an image to open the full-size original PNG (5472x3072, 16 MP). All originals are in the examples folder.

Tortoise island Teacup tsunami
Book canyon Moon hatching
Lightbulb orchard Mirror desert
Octopus pizza Bathroom mirror

The octopus pizza is 3840x2176 (8 MP); all others are 5472x3072 (16 MP).

Prompt guide: let any AI write your prompts

MINIMAX_H3_IMAGE_PROMPT_GUIDE.md teaches an AI chatbot how to write prompts for this model. Drop the file into ChatGPT, Claude, Gemini or a local model, then ask for what you want, for example:

  • "Using this guide, write me a prompt for a rainy Tokyo street at night with a woman holding a red umbrella."
  • "Using this guide, write an edit prompt: Picture 1 is my living room, make it sunset and add a cat on the sofa."
  • "Using this guide, put the man from Picture 1 into the café from Picture 2."

It answers with a finished prompt plus the settings to use (aspect ratio, draft or full size). Paste the prompt into the PROMPT box and leave Enhance prompt off, because the prompt is already long.

The guide covers text-to-image, single-picture edits and multi-reference scenes, plus realistic skin, hands, body direction, text in images, and fixes for common problems.

What is in the workflow

  • Prompt box with an optional Enhance prompt switch: a Qwen3.5 9B upsampler rewrites a short idea into a long, detailed caption (text-to-image only).
  • Aspect ratio picker (1:1, 3:4, 4:3, 2:3, 3:2, 9:16, 16:9, 21:9) and a full / draft switch. Use the draft size to test prompts quickly.
  • Ten reference picture slots for editing and multi-image prompts. They are all bypassed by default: select a slot, press Ctrl+B to switch it on, and refer to the pictures as <Picture 1>, <Picture 2>, ... in the prompt.
  • Sampler: SamplerLCM (s_noise 1.1 -> 0.9) + BasicScheduler simple, 6 steps, CFG 1.0, no negative prompt. The model is distilled for 6 steps, so more steps do not help.
  • A How to use note inside each workflow with the exact sizes for every aspect ratio.

Model files

All four come from the official Comfy-Org repo: Comfy-Org/MiniMax-H3-Image. Click a file name to download it, then put it in the folder shown inside your ComfyUI models folder.

Role File (direct download) Size Folder
Diffusion model minimax_h3_image_pruned_int8_convrot.safetensors 21 GB models/diffusion_models/
Text encoder qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors 15.7 GB models/text_encoders/
Prompt upsampler (optional, only for Enhance prompt. 8 GB: skip it, it is switched off in the 8 GB workflow because it makes images very slow) qwen3.5_9b_w4a8_gptq.safetensors 7.3 GB models/text_encoders/
VAE minimax_h3_image_vae_fp16.safetensors 1.2 GB models/vae/

About 38 GB in total, or 45 GB with the optional upsampler.

No custom nodes needed. Every node in these workflows is built into ComfyUI. Update ComfyUI to the latest version (tested on 0.45.3); older versions may not have the MiniMax H3 Image nodes.

Low VRAM tips

  • The model files are bigger than the card (about 21 GB diffusion + 16 GB text encoder). ComfyUI reads them in parts, so they do not need to fit in VRAM. Tested with 24 GB system RAM. A fast SSD helps the first load.
  • The 4 MP draft is handy for testing prompts. Switching to 16 MP with the same seed gives a different picture, not the same one bigger, so if you like a 4 MP image, keep it.
  • Reference pictures: all 10 slots work on every card, but each picture makes the steps slower. On 8 GB, 4 pictures take about 4 minutes at 16 MP; with many pictures, set Full megapixels to 8 to save time.
  • Close other apps that use the GPU (games, video in the browser, OBS preview).
  • If the decode runs out of memory, lower tile_size on the tiled VAE Decode node to 384 or 256.
  • Prompts: write long, photo-like descriptions (who, where, light, camera). Short prompts give a glossy, doll-like face.
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support