MiniMax H3 Image Review: Strengths, Weaknesses and the Best Settings

16 megapixels in seconds and edits with up to 16 reference images: those are the figures MiniMax used when it released H3 Image on 10 November 2026, the open image spin-off of its well-known video model MiniMax H3. Shortly before that, the Munich start-up Pruna AI released a second image model on the same base, P-H3-Image. This MiniMax H3 Image review covers what actually holds up in practice, where H3 Image loses to Qwen Image and Krea 2, and how you can get better results with a few settings.1,4
H3 Image is strong at editing: putting people into new rooms, swapping clothes, rewriting signs. Light and reflections adapt convincingly. It is weak on fine detail: backgrounds turn mushy when you zoom in, and faces don't always stay the same during edits. For images from text there are better open models, but for edits with many reference images H3 Image is among the best. In Germany, the easiest way to use it is the official API, from about 2 US cents per image. You can build matching prompts with the prompt builder further down.
What I noticed
- Editing is its star discipline. I put a person from a reference image onto an empty chair: she sits naturally in the room, her reflection appears on the monitor, and furniture and cables stay where they were.
- Faces are the weak spot when editing. Sometimes the person stays perfectly recognisable, sometimes a different face comes out. The facial expression tends to go neutral.
- Fine detail can't keep up with the 16 megapixels. The overall impression is right, but when you zoom in, fences, blossoms and backgrounds turn soft.
- It's fast. A 2K image takes a little over 10 seconds on a graphics card with 16 GB of VRAM.
- Large text works, small text doesn't. Signs and headlines are mostly spelt correctly, while small handwriting contains mistakes.
MiniMax H3 Image compared: is it the best open image model?
Not at everything. MiniMax itself describes H3 Image as the open model with the strongest instruction following.1 In the comparisons, the result depends heavily on the task:
| Task | Ahead | Assessment |
|---|---|---|
| Editing with several reference images | H3 Image | Places people and objects convincingly in space, up to 16 reference images |
| Preserving the face when editing | Qwen Image 2.1 | Keeps identity more reliably, H3 Image varies |
| Image quality from text | Krea 2 | Looks more natural and less over-smoothed; H3 Image often looks too perfect |
| Difficult poses | mixed | H3 Image manages a head sticking out of a car window better, but doesn't recognise the pigeon pose in yoga |
| High resolution and speed | H3 Image | 16 megapixels directly, without upscaling |
| Styles | mixed | H3 Image does magazine covers and user interfaces well, Krea 2 is better at pixel art and the Monet style |
Compared with the closed paid models GPT Image and Nano Banana, H3 Image scores mainly on speed and spatially coherent edits. One example: the reflection of an inserted person on the monitor was missing in the GPT Image comparison picture.
P-H3-Image is a serious alternative. It produces 8 instead of 16 megapixels, is designed for 8 steps and can be added to an existing H3 installation as an add-on module of around 2 GB. It does well on pixel-accurate edits: the image doesn't shift and the colours stay the same.4 Pruna itself names photorealism, film and editorial looks, and graphic styles with short text as strengths.4
The two models side by side:
| MiniMax H3 Image | P-H3-Image | |
|---|---|---|
| Provider | MiniMax | Pruna AI |
| Image size | around 16 megapixels, e.g. 4096 × 4096 | 8 megapixels for text-to-image, 4 or 8 for editing |
| Steps | 6 | 8 |
| Reference images when editing | up to 16 | up to 8 |
| Model size | 33.1 bn parameters, BF16 around 69 GB | 33.1 bn parameters, BF16 around 100 GB; as an add-on module for H3 around 2 GB plus text encoder and VAE |
| Prompt enhancer | optional: its own model H3-Image-PE, Qwen3.5-9B in ComfyUI | on by default, based on Qwen3.5-9B |
Strengths in detail
Editing with reference images
H3 Image and P-H3-Image handle generating and editing in the same model. You don't need a separate editing model.1,4 These things work well:
- Putting people into new scenes, even from a character sheet with front, side and back views
- Swapping clothes, including prints with text
- Replacing words on signs, in the same typeface
- Changing the time of day, for example a house from a grey day to a clear night
- Restoring and colourising old photos
- Printing logos onto products, including fabric texture
The reason for the spatial strength lies in its origin: H3 Image is based on the video model MiniMax H3 and takes over its architecture.1 A video model has to understand how rooms, light and reflections fit together, and that knowledge is also in the image model.
Text in images
H3 Image mostly writes large words in quotation marks without errors, including Chinese characters next to English text. A menu with over 50 words was completely correct at 8 megapixels, while at 1 megapixel the small handwriting fell apart.
Speed
| Graphics card | Model | Result |
|---|---|---|
| 16 GB VRAM | H3 Image, 4-bit (W4A8) | 2K image in approx. 11 s, edit with 4 MP in approx. 26 s |
| RTX 4090 | H3 Image | 16 MP in approx. 40 s |
| RTX 5090 | P-H3-Image, with prompt enhancer | 2 MP in approx. 30 s, 4 MP in approx. 1 min |
| 24 GB VRAM | P-H3-Image, without prompt enhancer | 8 MP in approx. 20 s |
The figures are rules of thumb from practical tests with ComfyUI, not a lab benchmark. Whether the prompt enhancer runs alongside makes a big difference. Low system memory slows things down a lot: with 32 instead of 64 GB, the computer reloads the models on every run, and an image then takes about four times as long. According to Pruna, P-H3-Image in full precision needs around 101 GB of graphics memory at 8 megapixels, while the 4-bit version for ComfyUI gets by with 21 GB.4,6
Weaknesses in detail
Faces when editing
If you put a person from a reference image into a new picture, a different face can come out, often enough that you should check every result carefully. A neutral facial expression is also typical. Capital letters in the prompt don't change that. What helped was describing the expression in simple words, such as “mouth slightly open, eyebrows raised”, and explicitly stating that the face and facial features stay unchanged.
Mushy details
There is a technical reason for the soft details. The image encoder shrinks the image by a factor of 16, and the model then merges 2 × 2 cells once more. Each compute unit therefore stands for 32 × 32 pixels.1 For a video model with many reference images that makes sense, but for fine image detail it is a drawback.
Recurring failure patterns
- Animals get human hands, such as a rabbit playing the guitar. Fix: explicitly ask for “paws” in the prompt.
- Unwanted logos, for example a manufacturer's name on a guitar although the prompt didn't ask for one.
- Poses shift during edits, for example crossed arms that disappear.
- Style transfer from a second image works poorly. The model doesn't recognise cubism, and oil painting stays almost a photo.
The best settings and prompt tips for MiniMax H3 Image
- Describe in detail, in English. According to the model card, H3 Image expects a detailed English description of the whole image. For short prompts, the MiniMax API first has an add-on model flesh them out.1
- Render large enough. MiniMax gives a working range of roughly 4 to 20 megapixels, with the best results at around 16.1 For quick drafts you can go lower, but small faces need at least 4 megapixels.
- Don't increase the steps. H3 Image is designed for 6 steps and P-H3-Image for 8, both without CFG and without a negative prompt.1,4 With 20 steps the image came out overcooked.
- Put text in quotation marks and leave out filler words like “masterpiece” or “8K”. They make the image glossy and oversaturated.
- When editing, say what stays. At the end of the prompt, explicitly state that the face, facial features, pose and framing remain unchanged. Describe the facial expression plainly instead of forcing it with capital letters, for example “mouth slightly open, eyebrows raised”. Only name what is really in the image, otherwise details bleed onto other areas.
- The first image sets the format. The result takes over the aspect ratio of the first image, so load the background first.
- When removing something, say what is there afterwards. “Remove the lettering” on its own often fills the spot with something new. “A smooth, dark green wooden surface without letters” works.
- Try samplers. With LCM instead of Euler, results often get softer and have fewer artefacts.
The built-in prompt enhancer turns a short idea into a long prompt and visibly improves images from text. But it needs several extra gigabytes of VRAM, can take a long time on weaker machines and, without precise instructions, invents its own details, such as a band name on a poster.
Prompt builder: editing prompts in the official format
In the official MiniMax examples, an editing prompt has a fixed structure. First, each object is assigned to an image in English, for example “The woman on the left in <Picture 1> is <Subject 1>”. Then comes what the target image shows and, for each object, what is preserved and what changes.7 Otherwise the prompt enhancer writes this structure.
With the builder you fill in the parts yourself and get the finished prompt. It warns you about the typical mistakes: filler words, capital letters, negations, animals without paws, and removing something without saying what should be visible afterwards.
Prompt builder for MiniMax H3 Image
Builds your prompt in the format MiniMax uses in the official examples. Fill in the fields in English, because the model expects an English description. Everything runs in your browser, nothing is sent.
Works via the MiniMax API, the Hailuo app and, where the licence allows it, locally. The prompts can usually be used for P-H3-Image and other image editors as well.
How to get started with MiniMax H3 Image: API or ComfyUI
Without installing anything, you can use the Hailuo web app or the API of the MiniMax platform. The API has separate endpoints for text-to-image, editing and prompt expansion.1,11 Billing is pay-as-you-go. MiniMax gives these example prices per image:13
| Task | Result | Price per image |
|---|---|---|
| Image from text | 1K (1024 × 1024) | approx. US$0.022 |
| Image from text | 2K (2048 × 2048) | approx. US$0.023 |
| Image from text | 4K (4096 × 4096) | approx. US$0.069 |
| Edit with 1 reference image | 4K | approx. US$0.073 |
| Edit with 3 reference images | 4K | approx. US$0.079 |
As a launch offer, 4K images are half price until 17 November 2026. A faster variant called H3-Image-Highspeed costs about double.13 You can build the API into workflows with n8n, Make or an AI agent. Bear in mind that your images go to MiniMax in the process. Customer photos belong there only with a legal basis.
In ComfyUI you need the current version. After that, under “Templates” you'll find one template for text-to-image and one for editing. The official ComfyUI files:5
| File | Size | Folder |
|---|---|---|
| Diffusion model (INT8) | 21 GB | models/diffusion_models |
| Text encoder Qwen3-VL-32B (INT8) | 25.3 GB | models/text_encoders |
| VAE (FP16) | 1.2 GB | models/vae |
| Prompt enhancer Qwen3.5-9B (optional) | 7.3 GB | models/text_encoders |
If you already work with the video model MiniMax H3, you usually already have the text encoder. For local use, though, the licence question in the next section applies.
Can I use MiniMax H3 Image locally in Germany?
This is the big question if you want to run the model on your own computer in ComfyUI. The H3 Image licence applies worldwide except for the EU, the United Kingdom, South Korea and the USA. That expressly includes the generated images.2 The same territorial rule applies to P-H3-Image.4 MiniMax justifies this with the still unclear rules for generative models, including the EU AI Act, and offers an application for a separate licence for these countries.3
According to MiniMax, the restriction does not apply to the official API.3 So you can also use H3 Image from Germany via the MiniMax platform, the Hailuo web app or OpenRouter.1,11,12 If you still rely on self-hosting for your image editing, check the licence text first for every open model. This is not legal advice.
Who is behind H3 Image and P-H3-Image?
MiniMax is an AI company from China. It develops its own foundation models for text, image, video, speech and music and is best known for the video app Hailuo. By its own account, more than 300 million people in over 230 countries and regions use its models and products.8 H3 Image is the image spin-off of the video model MiniMax H3, which came out in the summer.1
Pruna AI is a start-up from Munich with a second research site in Paris. Its founders include Stephan Günnemann, professor at the Technical University of Munich. Pruna makes AI models smaller, faster and cheaper to run.9,10 P-H3-Image is exactly that kind of project: Pruna turned the large H3 video model into an image model.4
Conclusion: who MiniMax H3 Image is worth it for
Worth it if you want to edit existing images: putting product photos into new settings, bringing together people from several reference images, changing captions or restoring old photos.
Probably not if you mainly need detailed, natural-looking images from text. For that, Krea 2 was mostly the better choice in the comparisons. And if faces have to stay exactly the same during edits, Qwen Image 2.1 is more reliable.
Frequently asked questions about MiniMax H3 Image
Is MiniMax H3 Image better than Qwen Image 2.1?
It depends on the task. H3 Image is ahead on high resolution, speed and spatially coherent edits. Qwen Image 2.1 was more reliable at preserving faces in the comparisons.
Why does MiniMax H3 Image change the face when editing?
It doesn't always happen, but it does regularly. What helps is at least 4 megapixels for small faces, a description of the expression in simple words and a line that explicitly states the face stays unchanged.
What is the difference between MiniMax H3 Image and P-H3-Image?
H3 Image comes from MiniMax itself, produces around 16 megapixels in 6 steps and takes up to 16 reference images. P-H3-Image comes from the Munich start-up Pruna AI, produces 8 megapixels in 8 steps and takes up to 8 reference images. Both are based on the video model MiniMax H3.1,4
How much graphics memory does MiniMax H3 Image need?
The full BF16 version is around 69 GB, plus the text encoder.1 For ComfyUI there is an INT8 version of 21 GB plus a 25.3 GB text encoder. The template deliberately starts at 4 megapixels so that machines with little VRAM don't crash.5 Thanks to automatic offloading to system memory, H3 Image also ran quickly on a card with 16 GB of VRAM.
How much does MiniMax H3 Image cost?
The model files are free, but according to the licence they may only be used outside the excluded territories, which include the EU.2 Via the API, an image costs between about US$0.022 (1K) and US$0.069 (4K) according to MiniMax, with edits a little more.13
How do I install MiniMax H3 Image in ComfyUI?
Update ComfyUI, open the MiniMax H3 Image template under “Templates” and put the four files in the folders given: diffusion model, text encoder, VAE and, optionally, the prompt enhancer.5 In the EU, local use is only allowed under the licence with separate permission from MiniMax.2,3
Can I create videos with MiniMax H3 Image?
No, H3 Image only produces still images. The sister model MiniMax H3 is responsible for videos. You can use an image created with H3 Image there as the start image.1
Sources
- MiniMax: model card MiniMax-H3-Image on Hugging Face, accessed 12 November 2026. huggingface.co.
- MiniMax: MiniMax-H3-Image Community License Agreement, accessed 12 November 2026. huggingface.co.
- MiniMax: Q&A About License (MiniMax-H3), accessed 12 November 2026. huggingface.co.
- Pruna AI: model card P-H3-Image on Hugging Face, accessed 12 November 2026. huggingface.co.
- ComfyUI documentation: MiniMax H3 Image. docs.comfy.org.
- ComfyUI documentation: P-H3-Image. docs.comfy.org.
- MiniMax: official editing example (revised_prompt.edit.txt) for the model card. huggingface.co.
- MiniMax: About MiniMax, accessed 12 November 2026. minimax.io.
- Pruna AI: About, accessed 12 November 2026. pruna.ai.
- Technische Universität München (1 October 2026): Pruna AI develops optimized, highly efficient AI models. tum.de.
- MiniMax platform: Image Generation Guide. platform.minimax.io.
- OpenRouter: MiniMax H3 Image, model page with prices. openrouter.ai.
- MiniMax platform: Pricing Overview, Image section, accessed 12 November 2026. platform.minimax.io.
Measured values are rules of thumb from practical tests and depend on hardware and settings. As of: 12 November 2026.