MiniMax-H3-Image 16-Megapixel Image Generation Model: 4GB+ VRAM

MiniMax-H3-Image 16-Megapixel Image Generation Model: 4GB+ VRAM

At a Glance

Item Value
Repository Comfy-Org/MiniMax-H3-Image
Family guide MiniMax-H3-Image guide (1 articles)
Publisher guide MiniMax: models and licenses
Published 2026-11-09
License minimax-h3-community-license-agreement
Formats safetensors
Source type Unverified (not confirmed by a primary source)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

It is reported that an image generation model “Comfy-Org/MiniMax-H3-Image", optimized and repackaged for ComfyUI, has been released. Please note, however, that this news is currently unverified information that has not been officially confirmed.

This model is a reconfiguration of the open-weight model “MiniMax-H3-Image" developed by MiniMax AI, making it easier to use in a ComfyUI environment. It is a unified model that integrates text-to-image generation capabilities with image editing (image editing / image-to-image) capabilities that take one or multiple reference images as input.

Specifications

The specifications of this model and the architectural details of the original model (MiniMaxAI/MiniMax-H3-Image) are as follows:

  • Parameter Count and Architecture:
  • Diffusion Transformer (DiT): Employs a single-stream structure with 33.1B (33.1 billion) parameters. It features a total of 50 layers, a hidden size of 5376, 56 attention heads, and 3D multimodal RoPE. The weights are in BF16 precision, while the embedding layers, timestep layers, and output layers are in FP32 precision. – Text Encoder: Directly adopts the Qwen3-VL text encoder (using the hidden states of the 50th layer). – Image VAE: Composed of a CNN encoder and a 24-layer 2D ViT decoder, featuring 16x spatial downsampling and 24 latent channels.
  • Output Specifications:
  • Capable of directly generating high-resolution images of approximately 16 megapixels (16 million pixels) without using an upscale. The effective generation range is reported to be from 4 megapixels to 20 megapixels. – Examples of supported aspect ratios and generation resolutions: – 1:1 ratio: 4096×4096 – 4:3 / 3:4 ratio: 4736×3552 / 3552×4736 – 3:2 / 2:3 ratio: 5024×3360 / 3360×5024 – 16:9 / 9:16 ratio: 5472×3072 / 3072×5472 – 21:9 ratio: 6272×2688
  • Recommended Inference Settings:
  • Number of inference steps (num_inference_steps): Supports rapid generation in 6 steps (6 NFEs) via distillation. – Guidance scale (guidance_scale): 1 (CFG / Classifier-Free Guidance is not used). – Negative prompt (negative_prompt): Not supported.
  • License Conditions:
  • The ComfyUI repackaged version is provided under minimax-h3-community-license-agreement, and the original model is provided under minimax-h3-image-community-license-agreement.

Performance and Quality

The model card for the original model highlights three main features regarding generation quality and performance. Note that quantitative benchmark comparison tables with other models are not included in the published materials.

First, it achieves 16-megapixel output while keeping computational costs low. It is explained that while it can generate extremely high-definition images compared to typical 1 to 4 megapixel generation models, it keeps the amount of calculation within a manageable range.

Second, it offers detailed rendering capability and realism. It is reported that beyond the characteristic textures seen in conventional generated images, images that reproduce complex details and nuances of the real world can be obtained.

Third, it provides strong instruction-following and image editing capabilities. It is claimed that it has high prompt fidelity and enables advanced image editing that can integrate composition and elements by inputting up to 16 reference images.

In addition, a case study comparing execution scripts using the local inference framework “SGLang" and output results from directly calling the official API with the same prompt and settings is shown, reporting that local environments can also output at the same quality as the official API.

Strengths and Use Cases

Note that this news is unverified information without official confirmation. The following performance and use cases are descriptions provided by the creators themselves in the model card of the original model (MiniMaxAI/MiniMax-H3-Image).

  • High-Resolution Image Generation: Reported to directly output approximately 16 megapixels (4096×4096 to 6272×2688) without an upscaler. Suited for use cases requiring detailed rendering.
  • Photorealistic Expression: The model card lists examples such as “Photorealistic photography" and “Surreal editorial photography."
  • Illustration and Concept Art: Examples of environment concepts, characters, creatures, props, and painting/illustration styles are shown.
  • Images with Text: Examples are introduced where logic puzzles in Chinese and English are correctly solved on the board.
  • Image Editing: Examples include conversion from a single reference image (from 3D rendering to a cinematic scene style, or from a game map to a first-person perspective) and composition of 10 reference images (an empty room and 9 pieces of furniture) into a single living room. Up to 16 reference images can be specified.

Regarding prompts, the model card states that it assumes “English descriptions detailing the entire image." As a means to expand short sentences, the MiniMax API’s H3-Image-Context-IR and local MiniMax-H3-Image-PE (optional) are suggested.

This Comfy-Org version repackages the original model’s files for ComfyUI. Only file names are mentioned in the card. What changes were targeted and how they affect image quality is unknown as it is not stated in the materials.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch)

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) Original precision 0.0GB 0.0GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — minimax-h3-community-license-agreement (Commercial use prohibited): MiniMax’s MiniMax H3 Community License. Use of the weights is limited to the EU, the UK, South Korea and the US (Sections II and V.4); elsewhere, including Japan, a separate license is required.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

This model’s license (minimax-h3-community-license-agreement) does not permit commercial use. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Distributed Files

Weight files published in Comfy-Org/MiniMax-H3-Image, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size ComfyUI folder
diffusion_models/minimax_h3_image_pruned_bf16.safetensors 40.23GB models/diffusion_models/
diffusion_models/minimax_h3_image_pruned_int8_convrot.safetensors 20.98GB models/diffusion_models/
diffusion_models/minimax_h3_image_pruned_w4a8.safetensors 12.55GB models/diffusion_models/
text_encoders/qwen3.5_9b_w4a8_gptq.safetensors 7.34GB models/text_encoders/
vae/minimax_h3_image_vae_fp16.safetensors 1.17GB models/vae/

The ComfyUI folder is shown only where the repository’s layout mirrors ComfyUI’s models/ directory.

How to Get It

The distribution format is safetensors. According to the Comfy-Org/MiniMax-H3-Image card, the following three files are placed in ComfyUI:

ComfyUI/models/diffusion_models/minimax_h3_image_pruned_int8_convrot.safetensors
ComfyUI/models/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
ComfyUI/models/vae/minimax_h3_image_vae_fp16.safetensors

The text encoder is shared with MiniMax H3. It can be obtained from the text_encoders folder of Comfy-Org/MiniMax-H3. License agreement is not required for acquisition (gated is false). However, the license is a proprietary community license. Conditions such as commercial use should be checked in the license text of each repository.

Aside from ComfyUI, the model card for the original model recommends SGLang, vLLM, and diffusers. ComfyUI T2I/Edit templates are also noted, but the link destinations cannot be read from the materials. Commands when using the original model are as follows:

hf download MiniMaxAI/MiniMax-H3-Image
hf download MiniMaxAI/MiniMax-H3 --include "text_encoder/*"
# optional prompt expander
hf download MiniMaxAI/MiniMax-H3-Image-PE

In SGLang, start with the following command for a single GPU:

sglang serve --model-path MiniMaxAI/MiniMax-H3-Image --host 0.0.0.0 --port 30000 --warmup-mode off

To split a single request across multiple GPUs, specify parameters like --num-gpus 4 --ulysses-degree 4.

Points to note regarding requests are as follows:

  • num_inference_steps is only allowed to be omitted or set to 6, and guidance_scale is only allowed to be omitted or set to 1. Other values will result in an HTTP 400.
  • negative_prompt is rejected.
  • For reproducibility, it is recommended to specify a seed and set generator_device to cpu.
  • The default for min_generation_area is 4194304 (2048×2048). size values smaller than this will be generated on a large canvas and then resized.

The above API specifications describe usage via SGLang. Whether the same constraints apply in ComfyUI cannot be confirmed from the materials.

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-11-14 unsloth GGUF unsloth/MiniMax-H3-Image-GGUF そのままの精度 0.0GB (fits in 4GB VRAM)

File sizes of each build:

  • Available builds in unsloth/MiniMax-H3-Image-GGUF: そのままの精度 0.0GB

In addition, 5 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources

This article contains unverified information. We will append an update note once it is confirmed by a primary source.