P-H3-Image: Pruna shrinks MiniMax-H3 and needs just 8 steps
Pruna AI has built a slimmed-down version of the MiniMax-H3 image model: 7× smaller, 5× faster, and it needs only 8 steps instead of 50. It generates images up to 8 megapixels and edits images using up to 8 inputs. The license, however, excludes the EU, the United Kingdom and the US.
What it's about
Pruna AI, a company that specializes in compressing and speeding up AI models, has released the image model P-H3-Image on Hugging Face. It was built from MiniMax-H3 and can generate images from text (text-to-image) and edit existing images. The weights are open, and a ComfyUI variant is available too. According to Pruna's post on Reddit, the model runs 5× faster (8 steps instead of 50) and is 7× smaller than H3. These are the vendor's own figures.
What the model can do
P-H3-Image renders 8-megapixel images (for example 2912 × 2912 or 3872 × 2176). Pruna says 2 to 16 MP are also possible. It works with 8 steps and without CFG (classifier-free guidance, a technique that ties images more closely to the prompt but doubles the compute time). There is no negative prompt either.
For editing, you can pass in one to eight images. Image 1 is modified, and the others serve as references ("Picture 2", "Picture 3" …). For example, you can put the hat from one image onto a person in another. According to the description, the model is particularly strong at photorealism, documentary and editorial photography, film stills, anime looks, ink and charcoal drawings, screen-print posters and graphic design with short text in quotation marks.
Prompt upsampler
The model is sensitive to prompts: detailed, specific descriptions work much better than short requests. That's why a prompt upsampler based on Qwen3.5-9B is built in by default. It turns a single sentence into a detailed prompt matched to the aspect ratio. For edits, it produces an instruction covering what stays and what changes. It adds 6 to 11 seconds. You can remove it, in which case your prompt is used verbatim.
The developers' tips for writing your own prompts: - Say what the image is in the first sentence ("a two-color risograph print"). - Describe a scene, including the viewpoint, main subject and light source. - Put any text in exact quotation marks, keep it short and end with "no other text". - Avoid hype words like "stunning", "epic", "8k" or "masterpiece". They lead to smooth, oversaturated images.
For edits, Pruna recommends this form: "Edit . . Keep exactly as it is." Without the list of what to preserve, the model often redraws the entire image.
Numbers and hardware
| Value | |
|---|---|
| Transformer | 33.1 billion parameters |
| Total size, bf16 | approx. 100 GB |
| GPU memory, 8 MP | approx. 101 GB with upsampler, 83 GB without |
| With auto-offloading | peak approx. 76 GB (one 80 GB H100) |
| Time on B200, 8 MP | approx. 9 s (without upsampler) |
| ComfyUI int8/4-bit | 21 GB VRAM |
According to Pruna, the "7× smaller" figure is relative to H3. The team says it reworked all components: prompt upsampler, text encoder, diffusion schedule, transformer and decoder.
What you need to try it
The full Diffusers version needs data-center hardware: about 101 GB of GPU memory (H200, B200) or an 80 GB card with offloading. The ComfyUI variant (PrunaAI/P-H3-Image-Comfy) is meant for ordinary home PCs. It comes with int8 and 4-bit files that, according to the vendor, run in 21 GB of VRAM, so a 24 GB card should do. You'll also need enough system RAM and several tens of GB of disk space. Apple devices are mentioned in the code via "mps", but the source gives no practical figures. Google Colab and Kaggle are linked as notebooks, but the large variant hardly fits there.
Limitations
- Sensitive to prompts: very long prompts and hype words degrade images, and very short ones are weak without the upsampler.
- Hands can go wrong, and small people in the background can be distorted.
- Small or dense text turns into pseudo-text, while short strings work well.
- At 2 and 4 MP there are more artifacts in faces, hands and details; at 1 MP faces look waxy.
- Still images only, no video or audio.
Our take
A large image model shrinking to 8 steps and a 24 GB card is interesting for local users. But figures like "5× faster" come from Pruna itself, and there are no independent tests or quality comparisons with H3. The license is the critical point: the MiniMax-H3 Community License excludes the EU, UK, South Korea and the US without a separate license and requires "Powered by MiniMax H3". For readers in Germany, that means that without their own agreement with MiniMax, use isn't covered by the license text. Read the license before you use it. The text encoder and the upsampler, by contrast, are under Apache 2.0.
Sources
- Reddit: lokale Modelle & Bild-KI huggingface.co
- Reddit: lokale Modelle & Bild-KI huggingface.co
Try it yourself: P-H3-Image
Explained step by step, no prior knowledge needed – on your own computer or free in your browser.
What can I do with it?
P-H3-Image generates high-resolution images (up to 8 megapixels) from text and edits images from a short instruction, even with up to 8 reference images. It needs only 8 steps, making it much faster than many models that take 30–50 steps. It's strong at photorealism, editorial and film looks, anime, drawings and graphic design with short text. Weaknesses are hands, small background figures and long running text. A built-in prompt upsampler turns short requests into good prompts. The source doesn't say how it compares with well-known alternatives.
Who is it worth it for?
Interesting for users with a 24 GB card and ComfyUI experience who want high-resolution images and image editing locally, and for teams with large GPUs. Because of the license, it's problematic for users in the EU, UK, US and South Korea without a separate MiniMax license.
Ideas: What you can do with it
Describe a screen-print or risograph poster with a short title in quotation marks. The model handles the text and graphic style well.
Upload a photo and change a color or the setting, for example "let it snow on the street". The list of unchanged elements keeps the rest stable.
Provide up to 8 portraits or objects as references and have them merged into one scene, for example for illustrations or game ideas.
Generate 8 MP images in an editorial or film look as print templates for wall decor, provided the license allows it.
Describe individual film frames or anime frames to visualize scenes for a short film or tabletop roleplaying.
What your computer needs
For home PCs, use the ComfyUI variant PrunaAI/P-H3-Image-Comfy with int8 or 4-bit (approx. 21 GB VRAM). The bf16 Diffusers version needs approx. 101 GB, or approx. 80 GB with offloading.
- bf16 at 8 MP: approx. 101 GB of GPU memory with upsampler, 83 GB without.
- With auto CPU offload it runs on an 80 GB card (H100), peaking at approx. 76 GB.
- ComfyUI files (int8, w4a8/4-bit) run in 21 GB of VRAM, according to the vendor.
- Disk space: approx. 100 GB for the bf16 weights.
- On a B200, an 8 MP image takes approx. 9 s; on smaller cards it takes significantly longer.
How do I find out what my computer has?
- Windows: Press Ctrl + Shift + Esc → Task Manager → Performance tab → click GPU on the left. At the bottom you'll see Dedicated GPU memory (that's your VRAM). You'll find system memory under Memory.
- Mac: Apple menu at the top left → About This Mac. You'll see your chip (e.g. M2) and memory (e.g. 16 GB). On Apple chips, graphics and processor share this memory.
- Linux: Type
nvidia-smiin the terminal (shows your graphics card and its memory) andfree -hfor system memory.
Guide
Google Colab (in your browser, no installation)
Google Colab is a free service from Google. It gives you a computer with a graphics card in your browser for a few hours. All you need is a Google account. For this model: Colab Pro/Pro+ only – needs an A100 with 80 GB or an H100.
- Download the notebook. Click the button and save the file.
- Open Colab. Go to colab.research.google.com and sign in with your Google account.
- Upload the notebook. In the window that appears, click Upload on the left (or choose File → Upload notebook at the top) and select the file
colab.ipynb. - Turn on the GPU. In the top menu, go to Runtime → Change runtime type → under Hardware accelerator, select T4 GPU (with Colab Pro: L4 or A100) → Save.
- Start everything. Runtime → Run all (keyboard shortcut Ctrl + F9). If you see the warning “This notebook was not authored by Google,” click Run anyway.
- Wait. The cells run one after another. A green checkmark appears to the left of each cell when it's done. Installation and download take about 37 minutes here.
- Try it out. The result appears below the last cell. To test something else, change the text in the cell and run it again with Shift + Enter.
- Clean up. When you're done: Runtime → Disconnect and delete runtime. This saves your free quota.
Good to know: The free GPU is limited (usually a few hours per day). After about 90 minutes of inactivity, Colab disconnects, and you'll have to start everything over.
Windows 10 / 11
This model runs on Python. Our script handles the setup automatically: it creates an isolated Python environment, installs the required packages (PyTorch, approx. 3 GB) and downloads the model (100 GB (about 2.2 hr at 100 Mbit/s)).
- Update your graphics driver (NVIDIA): install the latest driver. Without an NVIDIA card, it runs on the CPU – much more slowly.
- Install Python. Download Python 3.12 from python.org and run the installer. Important: At the bottom of the first window, tick Add python.exe to PATH, then click Install Now.
- Create a folder and download the files. Create a folder, e.g.
Documents\AI-Test, and save both files into it: - Open a terminal in the folder. Open the folder in File Explorer, right-click an empty spot → Open in Terminal.
- Run it. Paste this command and press Enter:
Installation and download are running now. The first time, this can take 10–30 minutes.powershell -ExecutionPolicy Bypass -File try_windows.ps1 - Result. The folder now contains the generated image
output.png. - Try your own input: Open
demo.pyin an editor, change the sample text, and save. Then run this in the terminal:.\.venv-p-h3-image\Scripts\Activate.ps1 python demo.py
macOS
Macs with an Apple chip (M1 to M5) are ideal. Older Intel Macs are very slow.
- Open Terminal. Press Cmd + Space, type Terminal, and press Enter.
- Install the developer tools (first time only; includes Python):
Click Install in the window. If you see "already installed," everything is already in place.xcode-select --install - Download the files – both end up in your Downloads folder:
- Run it:
The first time, installation and download take 10–30 minutes.cd ~/Downloads bash try_mac.sh - Result. The folder now contains the generated image
output.png.
Linux
- Check your GPU.
nvidia-smishould show your card. If it doesn't, install the driver on Ubuntu and restart:sudo ubuntu-drivers install - Install the Python basics (Ubuntu/Debian):
sudo apt update && sudo apt install -y python3-venv python3-pip git - Put the files in a folder and open a terminal there:
- Run it:
bash try_linux.sh - Result. The folder now contains the generated image
output.png.
Try it
A street photographer's photo: a baker in Hamburg opens her shop in the morning, soft light from the left, no text.A two-color risograph print of a lighthouse on the North Sea, poster with the text "NORDSEE" at the top, no other text.A still from a Japanese anime TV series: two kids on bicycles in summer rain.Edit: Make the pickup truck red. Keep the background and the light exactly as it is.Put the hat from Picture 2 on the woman in Picture 1.Tips for better results
- Render at 8 MP if possible. Lower resolutions show more errors in faces and hands.
- Put text in exact quotation marks and keep it short.
- Leave the upsampler on if you only write short prompts.
- Use the "prompt" output to see the prompt the upsampler wrote and learn from it.
If something goes wrong
License: The MiniMax-H3 license excludes the EU, UK, South Korea and the US without an extra license. Read it first.
Diffusers has to be installed from GitHub (pip install git+https://github.com/huggingface/diffusers).
8 MP needs huge GPUs. On smaller cards, use the ComfyUI variant.
Hype words like "8k" or "masterpiece" make results worse.
For edits, always state what should stay unchanged, or the image gets redrawn.
Side lengths are rounded to multiples of 32.
“out of memory,” a crash, or your computer freezes
The model is too big for your VRAM (video memory). Pick a smaller variant from the table above and close other programs, especially games and browsers with lots of tabs.
It runs, but extremely slowly
Your CPU is probably doing the work instead of your graphics card. In Ollama, the PROCESSOR column of ollama ps shows how much runs on the GPU. If it says "100% CPU", update your graphics driver or pick a smaller variant that fits entirely into VRAM (video memory).
“ollama” or “python” is not found
Close the terminal window and open it again so it recognizes the newly installed program. For Python on Windows: reinstall it and check Add python.exe to PATH.
Windows: "Running scripts is disabled on this system"
Run the script with exactly the command above: powershell -ExecutionPolicy Bypass -File try_windows.ps1. This allows execution for this one time only.
Error message when loading the model (version, "trust_remote_code", unknown architecture)
Brand-new models often need the latest software version. With Ollama: update Ollama. With Python: pip install -U transformers in the activated environment. The Model Card usually lists the exact version.
Tokenstrom generates the scripts and notebook automatically. Please take a quick look before running them – they're plain, readable text. Mind the model's license: custom license (see Model Card).
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.