MiniMax-H3-Image (Pruned DiT) — GGUF Quantizations

Quantized GGUF weights for MiniMax-H3-Image, converted directly from the structurally pruned BF16 master checkpoint provided by Comfy-Org (minimax_h3_image_pruned_bf16.safetensors).

MiniMax-H3-Image is a state-of-the-art 33.1B parameter single-stream Diffusion Transformer (DiT) designed for ultra-high-resolution text-to-image generation and multi-reference image editing (up to 16 MP / 4096 × 4096) in just 6 distilled sampling steps.

All GGUFs in this repository have sensitive layers (the 24-channel input patchifier video_patch_proj, the 6-step PDD output head final_layer, and the low-rank AdaLN modulation basis adaln_basis / adaln_t_table) strictly preserved in unquantized FP32 to eliminate quantization noise and grid artifacts.

Model Files & Quantization Overview

File Name	Size	Quant Type	Recommended Hardware / VRAM	Quality / Use Case
`MiniMax-H3-Image-Pruned-Q2_K.gguf`	6.88 GB	`Q2_K`	8 GB – 10 GB VRAM	Maximum memory savings. Fits easily on consumer laptops & entry GPUs.
`MiniMax-H3-Image-Pruned-Q3_K_M.gguf`	8.91 GB	`Q3_K_M`	12 GB VRAM	High efficiency with significantly lower quantization loss than 2-bit.
`MiniMax-H3-Image-Pruned-Q4_K_S.gguf`	11.60 GB	`Q4_K_S`	12 GB – 16 GB VRAM	Fast 4-bit small variant. Well-balanced footprint.
`MiniMax-H3-Image-Pruned-Q4_K_M.gguf`	11.60 GB	`Q4_K_M`	16 GB VRAM	Recommended sweet spot. Near-lossless visual quality for standard GPUs.
`MiniMax-H3-Image-Pruned-Q5_K_M.gguf`	14.10 GB	`Q5_K_M`	16 GB – 24 GB VRAM	Maximum fidelity. Indistinguishable from full BF16.

🖼️ Sample Generations (Q4_K_M)

Required Companion Files

MiniMax-H3-Image requires a dedicated 24-channel Image VAE and the Qwen3-VL-32B text encoder (reading layer 50 hidden states).

1. Required GGUF Loader Patch (`minimax_h3_image_gguf_fix.py`)

Because ComfyUI-GGUF (UnetLoaderGGUF) does not currently forward the .safetensors header metadata (pdd_sigmas) or restore 2D shapes on FP32-protected PDD output heads prior to model_detection.py, loading a MiniMax-H3-Image GGUF without this patch will raise missing 1 required positional argument: 'pdd_sigmas' or a 24 vs 144 channel mismatch error.

Download minimax_h3_image_gguf_fix.py from this repository and place it directly inside your ComfyUI/custom_nodes/ folder:
```
ComfyUI/custom_nodes/minimax_h3_image_gguf_fix.py
```
Restart ComfyUI. It will automatically inject the 6-step pdd_sigmas schedule and 24-channel latent configuration in memory without modifying any ComfyUI core files on disk.

2. Image VAE (24-Channel)

Official / Comfy-Org FP16 VAE: minimax_h3_image_vae_fp16.safetensors
(Place inside ComfyUI/models/vae/)

Note: The Image VAE uses 24 latent channels with a 2D ViT decoder. Do not use standard 4-channel SD/Flux VAEs or the 16-channel video VAE from earlier H3 releases.

3. Text Encoder (Qwen3-VL-32B)

You can choose between pre-quantized safetensors or standalone GGUF text encoders:

GGUF Text Encoders (by Abiray):
- Q4_K_M (~18.8 GB): qwen3vl_32b_minimax_h3-Q4_K_M.gguf
- Q3_K_M (~14.2 GB): Qwen3vl_32b_h3-Q3_K_M.gguf
Official / Comfy-Org Text Encoders:
- NVFP4 / AWQ / BF16: Available in the Comfy-Org Text Encoders directory (qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors or qwen3vl_32b_minimax_h3_int8_convrot.safetensors)

(Place inside ComfyUI/models/clip/ or ComfyUI/models/text_encoders/)

📋 Example Workflow

Download and import the ready-to-use ComfyUI workflow directly from this repository: 👉 image_minimax_h3_image_t2i_gguf.json

Installation & File Paths in ComfyUI

Ensure you have updated to ComfyUI v0.45.0+ and have ComfyUI-GGUF installed.

ComfyUI/
├── models/
│   ├── unet/ (or diffusion_models/)
│   │   └── MiniMax-H3-Image-Pruned-Q4_K_M.gguf
│   ├── vae/
│   │   └── minimax_h3_image_vae_fp16.safetensors
│   └── clip/ (or text_encoders/)
│       ├── qwen3vl_32b_minimax_h3-Q4_K_M.gguf
│       └── (or qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors)

Recommended Generation Parameters

MiniMax-H3-Image is a distilled Piecewise Decoupled Diffusion (PDD) model trained specifically for single-stream 6-step inference:

Sampler: euler
Scheduler: simple or sgm_uniform
Sampling Steps: 6 (the fused PDD heads are calibrated for exactly 6 steps)
CFG Scale: 1.0 (No CFG; do not use high CFG or negative prompt scaling)
Denoise: 1.0
Latent Dimensions: Native 24 channels (2 × 2 patchification, 32 × 32 pixels per token)
Resolution: 1024 × 1024 (1 MP), 2048 × 2048 (4 MP), or up to 4096 × 4096 (16 MP)

Prompting Guide

MiniMax-H3-Image responds best to dense, descriptive English descriptions that specify spatial layout, material properties, lighting conditions, and typography.

Example Prompt (Packaging & Typography)

The image is a cheerful toy-packaging poster built around a handmade plush fox floating above a summer meadow, drawn in a warm, fairly saturated palette: clear sky blue deepening toward the top, rich burnt-sienna and terracotta plush for the fox, warm cream, fresh sage green in the meadow, cornflower and cherry-red accents on the small props, and buttercup yellow in the lettering, with bright afternoon sun from the upper left and crisp short shadows. The fox is the hero, sewn from long-staple mohair-type plush in warm burnt sienna and hand-stuffed so its silhouette is slightly lumpy; the pile is long and irregular, single fibres catching the light along its outline as a bright hairy halo against the sky, with cream muzzle and sock paws. Across the top of the frame the word "MINIMAX" is inflatable and rounded in thick felted wool; directly beneath, smaller and centred, "H3" sits between two small dots. Along the bottom, a heavy bold sans-serif logotype reads "TEXT TO IMAGE" in saturated cornflower blue. All lettering is English only, with no Chinese characters.

Example Prompt (Photorealistic Macro Still Life)

A raw, high-resolution 35mm documentary portrait of an artisan ceramicist in her sunlit studio loft, shot on a Leica M11 with a 50mm f/1.4 lens. Bathed in cool morning daylight streaming from a tall steel-framed window on the left. Authentic, unretouched skin texture with visible pores, fine smile lines, and a faint dusting of dry white clay powder across her cheekbone. Her hands rest on a weathered pine workbench scattered with trimming tools and a half-empty stoneware coffee mug. Organic shallow depth of field, natural color grading, no plastic skin smoothing or HDR halos.

License & Usage

This model is subject to the MiniMax-H3-Image Community License Agreement.

Free for research, personal, and commercial creation within the terms of the Community License Agreement.
If your monthly/annual revenue exceeds the threshold specified in Section IV of the license, or if you provide hosted API services, refer to the commercial licensing terms at platform.minimax.io/license.

Credits & Acknowledgements

MiniMax AI for developing and releasing the weights of MiniMax-H3-Image.
city96 for the ComfyUI-GGUF loader suite.
Quantized by Abiray.

Downloads last month: -

GGUF

Model size

20B params

Architecture

wan

Hardware compatibility

2-bit

3-bit

4-bit

5-bit

Model tree for Abiray/MiniMax-H3-Image-GGUF

Base model

MiniMaxAI/MiniMax-H3

Finetuned

MiniMaxAI/MiniMax-H3-Image

Quantized

(3)

this model