How to Run MiniMax-H3-Image Locally

MiniMax's MiniMax-H3-Image, the 6-step image generation and editing model of the H3 family, as GGUF for ComfyUI and stable-diffusion.cpp: per-tensor layouts from a sensitivity scan of this model and an importance matrix from its BF16 weights, measured on one stand against Comfy-Org's own files and every other MiniMax-H3-Image GGUF we found. The renders and measurements are public.

Built by the team behind Atomic Chat, the app that runs open-weight models locally and offline.
The image model from 7.6 to 21.4 GB on disk for ComfyUI (Unet Loader (GGUF)) and stable-diffusion.cpp; the files load in stock ComfyUI-GGUF as they are. Of the five other MiniMax-H3-Image GGUF repos we found on 11 November, only one loads that way.
Closer to the original at the same size: AD-Q5_K (13.8 GB) is 16% closer to the BF16 picture than Abiray's Q5_K_M (14.1 GB), AD-Q3_K (8.7 GB) 10 to 20% closer than the three Q3_K_M files (8.8 to 8.9 GB), and our Q8_0 30% closer than Comfy-Org's int8_convrot. At Q4 the pictures are a tie; see the comparison.
No uncensored version: the MiniMax-H3-Image license forbids circumventing safety guardrails.

MiniMax-H3-Image is MiniMax's text-to-image and editing model of the MiniMax H3 family: one checkpoint for both, 6 steps, no CFG, 4 to 20 megapixels, edits with up to 16 reference images. Its diffusion transformer is H3's 33B design without the audio layers; with the AdaLN branches folded into a small table (Comfy-Org's pruned repack) it is 20.1B weights, 40.2 GB in BF16. This repo has that image model as GGUF. The text encoder (MiniMax-H3's Qwen3-VL-32B, unchanged) and the VAE stay the files you already use: Comfy-Org's for ComfyUI, or any H3 encoder GGUF for stable-diffusion.cpp.

Pick a file

Every file rendered the same 40 captions at 2048×2048 from the same starting noise as the BF16 weights. Two numbers per file, both against BF16:

Output error follows the BF16 path: at each of the 6 steps the file gets exactly the input the BF16 model got, and we measure how far its prediction lands (relative L2; 0 = identical). It is the stable number.
LPIPS is the picture you actually get, compared with the BF16 picture of the same caption and seed (lower is closer). vs. new seed puts it on a scale: 100% is how far BF16 itself moves when only the seed changes (LPIPS 0.627).

File	Size on disk	Bits per weight	Output error [95% CI]	LPIPS [95% CI]	vs. new seed
`BF16` (Comfy-Org's `minimax_h3_image_pruned_bf16.safetensors`, default attention)	40.23 GB	16	0	0	0%
`Q8_0`	21.45 GB	8.53	0.020 [0.017, 0.023]	0.090 [0.072, 0.111]	14%
`AD-Q6_K`	16.67 GB	6.63	0.033 [0.027, 0.039]	0.130 [0.105, 0.159]	21%
`AD-Q5_K`	13.84 GB	5.51	0.049 [0.041, 0.057]	0.164 [0.138, 0.191]	26%
`AD-Q4_K`	11.57 GB	4.60	0.076 [0.064, 0.088]	0.254 [0.224, 0.285]	41%
`AD-Q3_K`	8.73 GB	3.47	0.132 [0.117, 0.147]	0.334 [0.305, 0.366]	53%
`AD-Q2_K`	7.57 GB	3.01	0.170 [0.152, 0.187]	0.379 [0.349, 0.409]	60%

Q8_0 is the closest to the original. Pick it if it fits. It is closer to the BF16 picture than Comfy-Org's own BF16 file as ComfyUI ran it on our H100 (that file switches on int8 attention where ComfyUI offers it: LPIPS 0.109, ours 0.090, 32 of 40 captions closer), and 30% closer than Comfy-Org's int8_convrot (0.129) at 0.5 GB more.
AD-Q6_K matches Comfy-Org's int8_convrot picture for picture (0.130 vs 0.129) at 4.3 GB less.
AD-Q5_K and AD-Q4_K keep the composition of most pictures; small objects, faces and text strokes move.
AD-Q3_K and AD-Q2_K are for tight memory: more pictures change their layout (53% and 60% of a new seed); text and objects stay readable in the samples below. AD-Q2_K has a Q2_K base but 3.0 bits per weight: the sensitive last 20 blocks are at Q3_K-Q4_K.
The text encoder comes on top: Comfy-Org's int8 encoder is 27.1 GB on disk (nvfp4_awq 15.7 GB, unsloth's Q4_K_M GGUF 18.2 GB); ComfyUI moves the encoder and the image model in and out of VRAM as needed.

Compared with the other files

Comfy-Org's own smaller files and the GGUFs of the five other MiniMax-H3-Image GGUF repos we found on 11 November, on the same stand. Hearthscale's Q3_K_M, Q4_K_M and Q5_K_M are the plain rows of the next table (the same tensor data); realrebelai's other four files and Abiray's Q4_K_S were not measured. "Loads" means stock ComfyUI with ComfyUI-GGUF 6ea2651 (current main) and Unet Loader (GGUF), nothing patched. Files that do not load were measured through a copy with the metadata ComfyUI-GGUF needs written in (the architecture, and the config key where it was missing; tensor data unchanged byte for byte), so the quality columns show what the weights can do once a loader accepts them.

File	Size on disk	Loads as published	Output error	LPIPS [95% CI]	Closest of ours (LPIPS)
Comfy-Org `minimax_h3_image_pruned_int8_convrot.safetensors`	20.98 GB	yes	-	0.129 [0.108, 0.151]	`Q8_0` 21.45 GB: 0.090 [0.072, 0.111]
Comfy-Org `minimax_h3_image_pruned_w4a8.safetensors`	12.55 GB	yes	-	0.274 [0.242, 0.306]	`AD-Q4_K` 11.57 GB: 0.254 [0.224, 0.285]
Comfy-Org `minimax_h3_image_pruned_bf16.safetensors` as published (int8 attention)	40.23 GB	yes	-	0.109 [0.090, 0.129]	`Q8_0` 21.45 GB: 0.090 [0.072, 0.111]
Hearthscale `Q8_0`	21.45 GB	no: no metadata at all	-	0.098 [0.080, 0.117]	`Q8_0` 21.45 GB: 0.090 [0.072, 0.111]
LSDJesus `Q6_K`	17.52 GB	no: architecture `minimax_h3`, not yet accepted by ComfyUI-GGUF	-	0.137 [0.115, 0.160]	`AD-Q6_K` 16.67 GB: 0.130 [0.105, 0.159]
realrebelai `Q6_K`	16.74 GB	yes	-	0.131 [0.111, 0.151]	`AD-Q6_K` 16.67 GB: 0.130 [0.105, 0.159]
lingjiu09 `Q4_0`	14.22 GB	no: architecture `minimax_h3`, not yet accepted by ComfyUI-GGUF	-	0.295 [0.268, 0.323]	`AD-Q5_K` 13.84 GB: 0.164 [0.138, 0.191]
Abiray `Q5_K_M`	14.08 GB	no: no `config` key, so ComfyUI stops at a missing `pdd_sigmas`	-	0.195 [0.164, 0.232]	`AD-Q5_K` 13.84 GB: 0.164 [0.138, 0.191]
LSDJesus `Q4_K_imat`	12.55 GB	no: architecture `minimax_h3`, not yet accepted by ComfyUI-GGUF	0.090 [0.080, 0.100]	0.255 [0.227, 0.285]	`AD-Q4_K` 11.57 GB: 0.254 [0.224, 0.285]
Abiray `Q4_K_M`	11.57 GB	no: no `config` key, so ComfyUI stops at a missing `pdd_sigmas`	0.097 [0.083, 0.110]	0.247 [0.223, 0.273]	`AD-Q4_K` 11.57 GB: 0.254 [0.224, 0.285]
LSDJesus `Q3_K_imat`	9.99 GB	no: architecture `minimax_h3`, not yet accepted by ComfyUI-GGUF	-	0.363 [0.334, 0.394]	`AD-Q3_K` 8.73 GB: 0.334 [0.305, 0.366]
LSDJesus `Q2_K_mix_imat`	9.25 GB	no: architecture `minimax_h3`, not yet accepted by ComfyUI-GGUF	-	0.429 [0.401, 0.457]	`AD-Q3_K` 8.73 GB: 0.334 [0.305, 0.366]
Abiray `Q3_K_M`	8.91 GB	no: no `config` key, so ComfyUI stops at a missing `pdd_sigmas`	-	0.397 [0.366, 0.430]	`AD-Q3_K` 8.73 GB: 0.334 [0.305, 0.366]
lingjiu09 `Q3_K_M`	8.81 GB	no: architecture `minimax_h3`, not yet accepted by ComfyUI-GGUF	-	0.418 [0.391, 0.445]	`AD-Q3_K` 8.73 GB: 0.334 [0.305, 0.366]
Abiray `Q2_K`	6.88 GB	no: no `config` key, so ComfyUI stops at a missing `pdd_sigmas`	-	0.546 [0.527, 0.567]	`AD-Q2_K` 7.57 GB: 0.379 [0.349, 0.409]

Loads. Our files load in stock ComfyUI-GGUF as they are. Of the other files we tried only realrebelai's Q6_K does: Abiray's files lack the config key that tells ComfyUI the 6 PDD steps; LSDJesus's and lingjiu09's declare the architecture minimax_h3, which ComfyUI-GGUF does not accept yet (city96/ComfyUI-GGUF#493 is open); Hearthscale's carry no metadata. They may work in forks of the node or in stable-diffusion.cpp, which ignores the metadata.
Same size, closer picture (paired over the same 40 captions): AD-Q5_K vs Abiray's Q5_K_M 0.164 vs 0.195 (31 of 40 closer), AD-Q3_K vs Abiray's Q3_K_M 0.334 vs 0.397 (34 of 40) and vs lingjiu09's Q3_K_M 0.418 (37 of 40), AD-Q2_K vs Abiray's Q2_K 0.379 vs 0.546 (40 of 40, ours is 0.7 GB larger), AD-Q4_K vs Comfy-Org's w4a8 0.254 vs 0.274 at 1 GB less.
Where we do not win: at Q4 the free-running pictures are a tie with Abiray's Q4_K_M (0.247, same 11.57 GB) and LSDJesus's Q4_K_imat (0.255, 12.55 GB); our output error is 22% and 16% lower (16 of 16 and 15 of 16 captions), but a 6-step model redraws enough of a picture at Q4 that LPIPS cannot separate them. At Q6 realrebelai's and LSDJesus's Q6_K tie with AD-Q6_K, and Hearthscale's Q8_0 ties with ours.
LSDJesus's files also keep Comfy-Org's int8-attention entries, so they run int8 attention in ComfyUI; their numbers include that.

What the layouts change

The plain types are what stable-diffusion.cpp's converter writes for one type everywhere. The AD files keep the same base type and move bits by a sensitivity scan of this model: each tensor kind (attention qkv and output, the two MLP matrices) and each band of ten blocks was pushed to Q2_K alone, and a planner spent the bytes where the output error drops most per byte.

Size class	Plain sd.cpp type	Output error	LPIPS	AD file	Output error	LPIPS
Q6	`Q6_K` 16.60 GB	0.033 [0.029, 0.037]	-	`AD-Q6_K` 16.67 GB	0.033 [0.027, 0.039]	0.130 [0.105, 0.159]
Q5	`Q5_K` 13.93 GB	0.060 [0.050, 0.070]	0.195 [0.165, 0.226]	`AD-Q5_K` 13.84 GB	0.049 [0.041, 0.057]	0.164 [0.138, 0.191]
Q4	`Q4_K` 11.43 GB	0.093 [0.080, 0.105]	0.258 [0.230, 0.286]	`AD-Q4_K` 11.57 GB	0.076 [0.064, 0.088]	0.254 [0.224, 0.285]
Q3	`Q3_K` 8.77 GB	0.165 [0.145, 0.183]	0.372 [0.343, 0.400]	`AD-Q3_K` 8.73 GB	0.132 [0.117, 0.147]	0.334 [0.305, 0.366]
Q2	`Q2_K` 6.73 GB	0.265 [0.242, 0.287]	-	`AD-Q2_K` 7.57 GB	0.170 [0.152, 0.187]	0.379 [0.349, 0.409]

Hearthscale's Q3_K_M, Q4_K_M and Q5_K_M carry the same tensor data byte for byte as the plain Q3_K, Q4_K and Q5_K rows (stable-diffusion.cpp's converter without an importance matrix); they were not rendered again.
The importance matrix alone lowers the output error by 5% at Q3_K and not at all at Q4_K; the layout does the rest, as with P-H3-Image.
The published AD-Q4_K is row AD-Q4_K-c in the metrics (no tensor below Q3_K; 11.57 GB). The first planner row AD-Q4_K (11.37 GB, Q2_K floor) measured 0.262 / 0.079 and is not published.

In ComfyUI

ComfyUI with MiniMax-H3-Image support (Comfy-Org/ComfyUI#17655, 10 November; we measured with 4d64a8b) and ComfyUI-GGUF.
Image model: one of these files in models/unet/, loaded with Unet Loader (GGUF).
Text encoder and VAE: from Comfy-Org/MiniMax-H3-Image: text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors (or _nvfp4_awq) with Load CLIP, type minimax, and vae/minimax_h3_image_vae_fp16.safetensors with Load VAE.
Graph: the ComfyUI template MiniMax H3 Image (image_minimax_h3_image_t2i), with Unet Loader (GGUF) in place of Load Diffusion Model. Keep its prompt expander: the model expects a long, detailed English description of the whole image, and its 6-step schedule comes from the file.

The files carry general.architecture = wan (the value ComfyUI-GGUF accepts today; ComfyUI itself recognises the model from its tensors) and the checkpoint's config with the 6 PDD steps and sigmas, which ComfyUI needs to build the model.

In stable-diffusion.cpp

MiniMax-H3-Image support is in stable-diffusion.cpp master since leejet/stable-diffusion.cpp#2324 (10 November). Built from 32d516c, our AD-Q4_K renders at 2048×2048 with unsloth's H3 encoder GGUF and Comfy-Org's VAE:

sd-cli --diffusion-model MiniMax-H3-Image-AD-Q4_K.gguf \
  --llm qwen3vl_32b_minimax_h3-Q4_K_M.gguf \
  --vae minimax_h3_image_vae_fp16.safetensors \
  -p "$(cat caption.txt)" -W 2048 -H 2048 --steps 6 --cfg-scale 1 --sigmas minimax-h3-image --diffusion-fa -o out.png

The encoder is MiniMax-H3's Qwen3-VL-32B unchanged (the same files as in MiniMaxAI/MiniMax-H3, text_encoder/), so an H3 encoder GGUF you already have works, e.g. qwen3vl_32b_minimax_h3-Q4_K_M.gguf from unsloth/MiniMax-H3-GGUF. sd.cpp renders the prompt as given: write a full caption, or expand it first.

Atomic Chat support is in progress (app, catalog): the app needs its stable-diffusion.cpp engine at build 1088 or newer, which is being built now.

How it was measured

Engine: ComfyUI 4d64a8b + ComfyUI-GGUF 6ea2651 on NVIDIA H100 80 GB, PyTorch 2.9.1 + CUDA 12.8.
Graph: the Comfy-Org template without the prompt expander: MiniMax H3 Image conditioning at 2048×2048 (the model's default area), KSampler 6 steps, cfg 1, euler, simple, seed 42 (43 for the floor).
Reference: Comfy-Org's minimax_h3_image_pruned_bf16.safetensors with ComfyUI's default attention. That file also asks ComfyUI for int8 attention (50 small comfy_attention.config entries); we measured it as published too (row "int8 attention" above) and left those entries out of our GGUFs, which run full-precision attention.
Held constant: Comfy-Org's int8 text encoder and fp16 VAE, so the image model file is the only thing that changes.
Captions: 40 requests in five categories (photo, illustration, English text, Chinese text, several objects with attributes), each expanded once into a long English description by an LLM prompt expander (the same captions as our P-H3-Image measurements). The importance matrix used 12 other captions.
Output error: 16 of the captions × 6 steps, the BF16 inputs recorded once and replayed through each file.
Not measured: edits, other resolutions and samplers, speed, aesthetics or prompt adherence as such.
In the metrics repo: every render, the captions, the importance matrix, the scan, the per-tensor layout of every file and every log.

How these files were made

Comfy-Org's diffusion_models/minimax_h3_image_pruned_bf16.safetensors from Comfy-Org/MiniMax-H3-Image, its 50 attention-method entries left out.
Converted with sd-cli -M convert from stable-diffusion.cpp 32d516c with --tensor-type-rules and --imat-in; then general.architecture, general.quantization_version and config are written into the metadata (the tensors are byte for byte the converter's).
The importance matrix comes from ComfyUI itself: a hook on every linear layer of the BF16 model while it rendered the calibration captions, written in stable-diffusion.cpp's format.
Kept out of quantization in every file: the AdaLN table and projections, the norms and biases (BF16; ComfyUI-GGUF cannot read quantized 1-D tensors), the patch layer and the 6 PDD output heads (F32) and condition_proj (BF16).

License

These files are a modified MiniMax-H3-Image Work under the MiniMax-H3-Image Community License Agreement; see NOTICE. MiniMax-H3-Image is licensed under the MiniMax-H3-Image Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.

The license is limited to its Applicable Territory: worldwide excluding the European Union, the United Kingdom, the Republic of Korea and the United States (section I.5). Read it before you use these files.
Its Acceptable Use Policy (Exhibit A) applies to every use, including the bans on circumventing safety guardrails and on harming minors, and the duty to disclose machine-generated content in public.
A separate written authorization from MiniMax is needed above US$1M annualized revenue or to offer a hosted service, and commercial products must display "MiniMax-H3-Image" (section IV).
The base model is MiniMaxAI/MiniMax-H3-Image by MiniMax. Quantized by AtomicChat; not affiliated with or endorsed by MiniMax or Comfy-Org.