P-H3-Image Released: Pruna Turns MiniMax-H3 Into an 8-Step Open Image Model (21 GB VRAM, ComfyUI Native)

Pruna AI has released P-H3-Image, an open-weights text-to-image and image-editing model carved out of the MiniMax-H3 video model. The headline numbers are big: Pruna says it is about 7x smaller than H3, needs 8 steps instead of 50, renders a 4 MP image in 2.6 seconds on a B200 (vs 12.4 s for H3), and runs in about 21 GB of VRAM at 8 MP with the 4-bit weights. It also works in ComfyUI natively, with no custom nodes. Here is what was released, what hardware you need, and the licence catch you should read before you build anything on it.

What Is P-H3-Image?

MiniMax-H3 is a huge open video model. If you have followed our MiniMax H3 ComfyUI guide, you know the full package is around 143 GB and asks for serious hardware. Video models are also very good image models, because a single frame is just a still picture, but nobody wants to load a 143 GB video stack to make a poster.

Pruna, a company that specialises in compressing and speeding up models, took H3 apart and rebuilt each piece for still images:

• Prompt upsampler: a Qwen3.5-9B based model that expands short prompts. It is on by default in Diffusers.
• Text encoder: cut from 32B to 4B parameters (Qwen3-VL-4B based), which drops it from 67 GB to 9 GB.
• Transformer backbone: cut from 33B to 20B parameters.
• Schedule: DMD distillation takes sampling from 50 steps to 8.
• Decoder: reworked for stills from 2 to 16 megapixels.

Speed and VRAM Numbers

These are Pruna’s own measurements. I have not benchmarked the model myself yet, so treat them as vendor numbers until community tests on consumer GPUs show up.

The important line for home users is the 21 GB figure. That points at 24 GB cards (RTX 3090, 4090, 5090) running the 4-bit build at full resolution. On a 16 GB card such as an RTX 5080 you will probably need lower resolutions, CPU offload, or a smaller quant from the community, so expect slower runs there. Note also that the Diffusers pipeline with the prompt upsampler in bf16 at 8 MP needs about 100 GB of GPU memory, so in Diffusers you will want the quantized weights or the upsampler turned off.

Every File, Size and ComfyUI Folder

The ComfyUI repository ships three complete sets. Pick one set, download all four files from it, and put each in the folder shown. Paths assume a standard install, so ComfyUI/models/diffusion_models/ means the diffusion_models folder inside your ComfyUI models folder.

Two notes. The w4a8 set has no VAE of its own, so it uses the int8 VAE. And if you already run MiniMax-H3, there is a 2.0 GB p-h3-image_lora_bf16.safetensors for ComfyUI/models/loras/ that turns your existing H3 install into an image model without downloading a new backbone. Pruna recommends the prompt upsampler in bf16 when it fits, and the smaller upsampler files only when 21 GB is your limit.

Image Editing and References

P-H3-Image is not only text-to-image. It can edit an input image, and it accepts several reference images to combine into one scene. Pruna’s own post is a little inconsistent on the limit (it mentions up to 8 and up to 9 references in different places), so start with fewer references and test. This puts it in the same space as Qwen-Image 2.1 Edit for multi-image edits.

Pruna is also open about the weak spots: very short, very long, or hype-heavy prompts give worse results; hands and small figures can distort, especially below 8 MP; and small or dense text is unreliable, while short quoted text works well. Write plain, descriptive prompts of medium length.

How to Get It in ComfyUI

Pruna says ComfyUI supports the model natively, so update ComfyUI to the latest version first and you do not need a custom node pack. The ComfyUI-ready weights are in the PrunaAI/P-H3-Image-Comfy repository, and the Diffusers weights are in PrunaAI/p-h3-image. Follow the file list on the model card for the exact folders. In a standard install, diffusion weights go in ComfyUI/models/diffusion_models/, text encoders in ComfyUI/models/text_encoders/ and the VAE in ComfyUI/models/vae/. If you only want to try it, there is a free playground Space at PrunaAI/raw-reindeer.

Which precision to pick:

• 4-bit (20 GB): 24 GB cards, or 16 GB cards with offload.
• 8-bit: 32-48 GB cards when you want closer-to-bf16 quality.
• bf16 (55 GB): data-centre GPUs, or for making your own quants.

Step-by-Step ComfyUI Setup and Settings

At release, P-H3-Image support is on ComfyUI’s master branch, and it ships to the desktop app and Comfy Cloud with the next stable release. So for now you need the latest ComfyUI from GitHub:

1. Update ComfyUI. In a terminal: cd ComfyUI then git pull. Portable Windows users can run update/update_comfyui.bat. Restart ComfyUI afterwards. If you get stuck at a “Reconnecting” message after updating, see our ComfyUI Reconnecting fix.
2. Download one file set from the table above into the listed folders.
3. Load an example workflow. The repo’s workflows/ folder has p-h3-image_text_to_image.json, p-h3-image_edit.json, p-h3-image_lora_text_to_image.json and p-h3-image_lora_edit.json. Drag the JSON onto the ComfyUI canvas.
4. Point the loaders at your files: Load Diffusion Model (the backbone), Load CLIP with type minimax (the text encoder), Load VAE, and the Generate Text node (the prompt upsampler).
5. Keep the sampler settings Pruna ships (table below) and press Queue.

Not sure what a sampler or scheduler does? Our ComfyUI samplers and schedulers explained guide covers it. Because this is a distilled 8-step model, do not raise the step count or add CFG expecting better quality; that usually makes distilled models worse.

Best Resolutions for Each Aspect Ratio

The model is tuned for 8 MP text-to-image and 4 MP edits. Sides are rounded to multiples of 32. Use these exact sizes in the Resolution Selector node:

Below 8 MP, Pruna warns that hands and small figures distort more, so if VRAM forces you to go smaller, keep people large in the frame.

Run It With Diffusers (Python)

If you prefer code to nodes, P-H3-Image ships as a Modular Diffusers pipeline. Install the latest Diffusers from GitHub:

pip install git+https://github.com/huggingface/diffusers
pip install -U transformers accelerate torch pillow

import torch
from diffusers import ModularPipeline

pipe = ModularPipeline.from_pretrained("PrunaAI/P-H3-Image")
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")

image = pipe(prompt="a lighthouse on a cliff at golden hour, film photo",
             height=3360, width=2496, output="images")[0]
image.save("lighthouse.png")

Memory in Diffusers is much higher than in ComfyUI: about 101 GB at 8 MP in bf16 with the upsampler, about 83 GB without it, and about 76 GB peak on one 80 GB GPU with automatic CPU offloading. On a B200 an 8 MP image takes about 9 seconds without the upsampler, and the upsampler adds 6 to 11 seconds. For home GPUs, ComfyUI with the int8 or w4a8 files is the practical route.

Editing in Diffusers: pass one image to edit, or a list of up to 8. Picture 1 is the one edited; the others are references named “Picture 2”, “Picture 3” and so on. Without the upsampler, Pruna suggests this template: “Edit <Picture 1>. <The change, described as the finished result>. Keep <everything that must stay> exactly as it is.”

Prompt Tips

• Describe the scene plainly: subject, setting, light, camera or style. Medium-length prompts work best.
• Skip hype words such as “8k” and “masterpiece”; Pruna says they push images towards an oversaturated look.
• For text in the image, keep it short and put it in quotes, for example a poster that says “OPEN LATE”.
• Leave the prompt upsampler on for short prompts. Turn it off when you have written a detailed prompt and want exact control.
• For edits, describe the finished result (“the jacket is red”) rather than the action (“make the jacket red”).

P-H3-Image vs Other Local Image Models

For setup guides on the alternatives, see Qwen-Image 2.1 Edit in ComfyUI and FLUX 3 Image Klein in ComfyUI.

The Licence Catch

This is the part most early posts skip. P-H3-Image inherits the MiniMax H3 Community License. According to Pruna’s announcement, that licence excludes the EU, the UK, South Korea and the United States unless you have a separate licence from MiniMax, and products built on it must show “Powered by MiniMax H3”. The Qwen-based text encoder and upsampler parts are Apache 2.0, and the ComfyUI nodes are MIT, but the main model weights follow the H3 terms. If you live in one of those regions or plan commercial use, read the licence on Hugging Face before you use it.

My Take

This is a smart release. H3’s image quality was always strong, but its size kept it out of reach for most people. Cutting the text encoder from 32B to 4B is the biggest practical win, because that encoder alone used to eat more memory than most consumer GPUs have. If community quants bring the 4-bit build comfortably under 16 GB, P-H3-Image could become one of the best local image models for 24 GB owners and a serious option for 16 GB cards. The licence, though, will keep many Western users and businesses on FLUX 3, Qwen-Image 2.1 and similar models with cleaner terms. For a wider comparison, see our best open-source text-to-image models roundup.

Frequently Asked Questions

Is P-H3-Image free?

The weights are free to download from Hugging Face, but use is governed by the MiniMax H3 Community License, which excludes several regions and requires attribution.

Can I run P-H3-Image on 16 GB VRAM?

Pruna lists about 21 GB peak at 8 MP with 4-bit weights. On 16 GB you will likely need lower resolutions or CPU offload. This is an estimate from the vendor figures, not a test.

Does P-H3-Image need custom nodes in ComfyUI?

No. Pruna says it is supported natively. Update ComfyUI and use the weights from PrunaAI/P-H3-Image-Comfy.

Is it the same as MiniMax-H3?

No. It is built from H3, but with a smaller text encoder, a smaller backbone, a new decoder and 8-step distillation. It makes still images and edits, not video.

Which P-H3-Image file should I download for a 24 GB GPU?

The w4a8 set: p-h3-image_w4a8.safetensors, the w4a8 text encoder and upsampler, and the int8 VAE. Pruna lists about 21 GB peak VRAM for it at 8 MP.

Why does ComfyUI not recognise P-H3-Image?

At launch, support is only on ComfyUI’s master branch. Update with git pull in your ComfyUI folder (or the portable update script) and restart. The desktop app gets it in the next stable release.

Can I use my existing MiniMax-H3 download?

Yes. Pruna provides a 2.0 GB p-h3-image_lora_bf16.safetensors for ComfyUI/models/loras/ plus LoRA workflows, so H3 owners can make images without a new backbone.

What settings does P-H3-Image use?

8 steps, euler sampler, no CFG and no negative prompt, both model sampling shifts at 1.0. Keep the sigmas from Pruna’s example workflow.

Sources

• Pruna AI: P-H3-Image launch post
• Hugging Face: PrunaAI/p-h3-image
• Hugging Face: PrunaAI/P-H3-Image-Comfy
• Hugging Face: MiniMaxAI/MiniMax-H3