Minimax H3 Image Quants - comfy-native - fp8_scaled / nvfp4 / GGUF Q6 / Q4 / Q3 / Q2_mix

作者：Thasmodius

MiniMax-H3-Image for ComfyUI: fp8, NVFP4 and GGUF The precisions of MiniMax-H3-Image that Comfy-Org's ComfyUI repack doesn't include. These files are the diffusion model only. The text encoder, the VAE and the bf16, int8 ConvRot and w4a8 versions are at https://huggingface.co/Comfy-Org/MiniMax-H3-Image Versions

• fp8 scaled, 21.0 GB: RTX 40 and RTX 50
• NVFP4, 12.5 GB: RTX 50, the fastest
• GGUF Q6_K 17.5 GB, Q4_K 12.5 GB, Q3_K 10.0 GB: any GPU, through ComfyUI-GGUF
• GGUF Q2_K mix, 9.25 GB: experimental, visibly softer Requirements
• ComfyUI 0.44 or newer (MiniMax-H3-Image support was added on 10 November 2026)
• Text encoder qwen3vl_32b_minimax_h3_nvfp4_awq (or int8convrot) and VAE minimax_h3_image_vae_fp16, both from Comfy-Org's repository above
• GGUF only: ComfyUI-GGUF must accept the "minimax_h3" architecture, or it stops with "Unexpected architecture type in GGUF file: 'minimax_h3'". Until that is merged upstream, run the patch script from the GGUF repo with the Python that runs your ComfyUI, then restart ComfyUI: https://huggingface.co/LSDJesus/MiniMax-H3-Image-GGUF/blob/main/patch_comfyui_gguf.py python patch_comfyui_gguf.py PATH/TO/ComfyUI python_embeded\python.exe patch_comfyui_gguf.py ComfyUI (Windows portable) Or make the edit by hand: add "minimax_h3" to IMG_ARCH_LIST near the top of custom_nodes/ComfyUI-GGUF/loader.py (the folder may be called comfyui-gguf). How to use Open ComfyUI's "MiniMax H3 Image" template and choose the file in Load Diffusion Model. For a GGUF, replace that node with Unet Loader (GGUF). Keep the template's settings: 6 steps, euler, simple, cfg 1. The model works best between 4 and 20 megapixels, does poorly at low resolution, and expects long, detailed English prompts (the template includes a prompt expander). What was changed Each file quantizes the same 200 transformer layers that Comfy-Org's int8 and w4a8 files do, 95.8% of the weights. Everything else is unchanged from Comfy-Org's bf16 file. Side-by-side comparisons and full details: https://huggingface.co/LSDJesus/MiniMax-H3-Image-fp8-nvfp4 https://huggingface.co/LSDJesus/MiniMax-H3-Image-GGUF License These files are modified (quantized) versions of MiniMax-H3-Image. MiniMax-H3-Image is licensed under the MiniMax-H3-Image Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved. Full license: https://huggingface.co/MiniMaxAI/MiniMax-H3-Image/blob/main/LICENSE By downloading or using these files you agree to that license, including its use restrictions (Section V) and its Acceptable Use Policy (Exhibit A), which apply to you. In particular:
• The license grants rights only outside the European Union, the United Kingdom, South Korea and the United States.
• The model and its outputs may not be used to improve any other AI model.
• Above USD 1 million in annual revenue, or to offer the model as a hosted service, you need written permission from MiniMax first: https://platform.minimax.io/license
• Commercial products and services using it must display "MiniMax-H3-Image". Powered by MiniMax-H3-Image. ==================================================================== ABOUT THIS VERSION (one per file) ================================= --- fp8 scaled (minimax_h3_image_pruned_fp8_scaled.safetensors, 21.0 GB) --- fp8 e4m3 weights with one scale per tensor. Runs natively on RTX 40 and RTX 50: 2.1 steps/s at 2048x2048 on an RTX PRO 6000. The MLP output layer (mlp.fc2) runs its matrix multiply in bf16, as in Comfy-Org's own H3 fp8 files; without that, images drifted badly. Modified file: quantized from Comfy-Org's bf16. --- NVFP4 (minimax_h3_image_pruned_nvfp4.safetensors, 12.5 GB) --- 4-bit floating point with an fp8 scale per 16 values. Runs natively on RTX 50 / Blackwell, and is the fastest version: 3.4 steps/s at 2048x2048 on an RTX PRO 6000. Other cards run it more slowly. Modified file: quantized from Comfy-Org's bf16. --- GGUF Q6_K (minimax_h3_image_pruned_Q6_K.gguf, 17.5 GB) --- 6.6 bits per weight; the closest of the GGUFs to the 8-bit versions. Needs ComfyUI-GGUF patched for minimax_h3 (see the description). GGUF is about 1 step/s at 2048x2048 on an RTX PRO 6000. If your card runs Comfy-Org's w4a8 (RTX 30 or newer) or int8 ConvRot (RTX 20 or newer), those are faster at a similar size. Modified file: quantized from Comfy-Org's bf16. --- GGUF Q4_K (minimax_h3_image_pruned_Q4_K_imat.gguf, 12.5 GB) --- 4.5 bits per weight, quantized with an importance matrix from 12 calibration images. Needs ComfyUI-GGUF patched for minimax_h3 (see the description). Modified file: quantized from Comfy-Org's bf16. --- GGUF Q3_K (minimax_h3_image_pruned_Q3_K_imat.gguf, 10.0 GB) --- 3.4 bits per weight with an importance matrix. The smallest version that holds up well. Needs ComfyUI-GGUF patched for minimax_h3 (see the description). Modified file: quantized from Comfy-Org's bf16. --- GGUF Q2_K mix, experimental (minimax_h3_image_pruned_Q2_K_mix_imat.gguf, 9.25 GB) --- Mostly 2-bit with an importance matrix. The first and last four blocks are at Q4_K, and the attention output and MLP output layers are at Q3_K, because plain 2-bit turns this model's images to noise. Images stay coherent but are visibly softer: fine texture smears and small text can break. Only 0.75 GB smaller than Q3_K. Needs ComfyUI-GGUF patched for minimax_h3 (see the description). Modified file: quantized from Comfy-Org's bf16.