Skip to content

P-H3-Image Edits at 8 Megapixels on a 24 GB Card, After You Read the Licence

Pruna rewrote MiniMax H3 into an 8-step image editor that renders 8 MP edits on quantized files in about 21 GB of VRAM and rewrites a one-line request before it draws. The three things the Japanese tutorial skips: the memory bill, the edit prompt form that keeps the rest of the picture intact, and a community licence that excludes the US, UK, EU and South Korea.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
A dark studio table with a large photographic print held under a single lamp, one corner of the print lifted to show a second surface beneath it
Editorial illustration about editing one region while the rest of the image stays put; not a model output.

The Japanese walkthrough opens with a complaint anyone who has retouched a photo with a generative model will recognise: you ask for one small change and the model rebuilds the whole picture. Then it makes a promise that should make you squint. Eight-step, 8-megapixel editing, on your own machine, for free, from one line of Japanese.

Both halves of that are real. P-H3-Image is Pruna AI's image-only rewrite of MiniMax H3, published to Hugging Face on 27 October, with roughly 398,000 downloads on the Comfy packaging by 6 November. The tutorial's workflow file sits behind the author's membership, which is fair enough. Two things it does not spend words on, and both decide whether this model belongs in your work: the memory bill, and the licence.

What the model actually is

MiniMax H3 is a video and audio model, and the Japanese post, citing Pruna's own numbers, puts its full weights at 143 GB. P-H3-Image keeps the parts that matter for stills: a 33.1B transformer, a 4B text encoder, a VAE, and a prompt upsampler, about 100 GB on disk in bf16.

Sampling is 8 steps with no classifier-free guidance, no negative prompt and no guidance scale. That is a shorter list of dials than most open image models hand you, and it tells you something about the design: most of the steering happens in the prompt, before the sampler runs. It renders text-to-image at 8 MP, edits one to eight input images at 4 or 8 MP, and makes stills only, no video and no audio.

The honest version of free

At 8 MP in bf16, the pipeline wants about 101 GB of GPU memory with the upsampler in the graph and 83 GB without, so the comfortable path is an H200 or a B200. Auto CPU offload brings the peak down to roughly 76 GB on an 80 GB H100. The Comfy packaging is the route for everyone else, and it splits into three sets rather than one: bf16 peaks at about 56 GB, int8 at about 31 GB, and w4a8 with an int8 VAE at about 21 GB. That last figure is the one that matters if you own a 24 GB card. Pruna reports that int8 stays close to bf16, often almost pixel for pixel, while w4a8 holds quality but moves more of the detail and sometimes the composition.

Free, in other words, means no per-image bill. It does not mean no hardware bill, and it certainly does not mean no electricity bill. On a B200 a text-to-image render takes about 9 seconds at 8 MP, with the upsampler adding 6 to 11 seconds on top. The tutorial's one to two minutes per image is a consumer-card figure running quantized weights, and that trade is the whole point: a minute per edit on hardware you already have is a different activity from waiting in an API queue, and it is why the local path exists at all.

Why 8 megapixels is not a spec-sheet number

The model card includes a resolution ladder rendered from one prompt and one seed, from 1 to 16 MP, with the face enlarged underneath. At 1 MP the face is smooth and waxy. By 8 MP, skin texture, stray hairs and small jewellery come through. That is what the memory buys: not a bigger canvas, but room for the sampler to place detail.

Two consequences follow. Renders at 2 and 4 MP break more often in faces, hands and small details, so shrinking the canvas to save VRAM costs you anatomy, not just size. And composition moves between resolutions, because the noise is tied to image size the way it is with any diffusion sampler, which is the mechanism behind why the same seed does not give the same picture at a different canvas size. Compare an 8 MP render against a 2 MP render and you are looking at two different images, not two different prompts.

The sizes Pruna documents for text-to-image: 2912 × 2912 for square, 3872 × 2176 for 16:9, 2176 × 3872 for vertical. Edits work best at 4 or 8 MP. Pass no height and width and the edit comes back at Picture 1's size, scaled up to at least 4 MP, which is usually the behaviour you want when the point is to preserve the original frame.

The rewriter is the real control surface

P-H3-Image runs a prompt upsampler, based on Qwen3.5-9B, on every request. It writes a detailed prompt for the aspect ratio you asked for, or, for an edit, a brief naming what to keep and what to change. Then it returns that written prompt as an output alongside the image.

Read it, and keep it. That string is the text the picture was actually made from, and for a one-line request it is the only reproducible part. A full record also needs the seed, the canvas size, the sampler settings and, for an edit, the source and reference images. ComfyUI's Save Image writes the prompt you submitted and the workflow into the PNG metadata, so those usually travel with the file; keep in mind that the metadata carries your original nine words rather than the rewritten prompt, so save that one yourself even when metadata is on, and if you strip metadata for delivery, copy the values out first. The upsampler runs with greedy decoding and caches its output, so the same request and the same images give back the same rewritten prompt instead of drifting between runs. If the render is wrong, the diagnosed problem is almost always in the rewritten prompt, not in your nine words.

Pruna's advice is to leave the upsampler in the graph, and that short prompts give weak results without it. That is the opposite of how most models behave, where the prompt is the prompt and any intermediate step is suspect. The Japanese workflow adds a second rewriter in front: a Gemma 4 12B vision model turns a short Japanese instruction into a fuller prompt before the built-in upsampler sees it. For a Japanese-speaking audience that is a reasonable layer, since it lets you write in the language you think in. It also costs you another model to load, another place for a bug to live, and a rewrite you cannot inspect unless you print it. If you are comfortable writing English prompts, the built-in upsampler already does this job.

Write the edit so the rest of the picture survives

The form Pruna documents is short enough to memorise:

Edit <Picture 1>. <the finished result>. Keep <what must not change> exactly as it is.

Four habits make it work. Name the exact object and where it sits: "the old man's faded blue cap" rather than "his hat". Describe the finished state instead of the action, so "the cap is now dark green wool" beats "change the cap colour". List what must stay, meaning faces, pose, background and light, because without that list the model tends to repaint the scene, which is the exact failure the tutorial opens with. And quote any new text exactly, naming the text around it as kept.

References go in the same request. Up to eight images, Picture 1 is the one being edited, the rest are addressable as Picture 2, Picture 3 and so on, so "give the woman in picture 1 the hat from picture 2" is a complete instruction. Size hints beat adjectives: "about the size of the cat's head" is a measurement, "a bit bigger" is a wish. That instinct, spell out what stays before you describe what changes, is the same one behind our guide to targeted edits.

Words that make it worse

Pruna's prompting notes are blunt in a way model cards rarely are: stunning, epic, cinematic, vibrant, hyper-detailed, 8k and masterpiece push the model toward a glossy, oversaturated, airbrushed look. Typing "8k" at a model that already renders 8 megapixels buys you plastic skin.

Open with what the picture is instead: a photograph by an experienced street photographer, a frame from a Japanese anime TV series, a two-colour risograph print, a flat scan of a handwritten recipe card. Then one scene, one viewpoint, one subject caught mid-moment, plus a few supporting elements and light with a source and a direction.

Styles the model handles well: documentary and editorial photography, portraits with real skin texture, film frames, TV-frame anime, ink and charcoal drawing, screen-printed posters, flat graphic design with short quoted text. Text past a few words turns into pseudo-text, so keep each string short, quote it, say where it sits, and close with "no other text". If you want no text at all, do not mention signs or labels, since naming them invites them in.

Limits, and the neighbours worth using instead

Hands still break, and small figures in the background can come out distorted. Rendering at 8 MP reduces both rather than fixing them. Dense text is unreliable. It makes single stills, full stop.

The Japanese post also positions it against two neighbours, and the ranking is worth repeating with the caveat that it is the author's judgment rather than a benchmark: MOSS Image Edit remains the strongest for pinpoint edits, Qwen-Image 2.1 is still the choice when you want a natural restyle, and P-H3-Image is the one to try first for broad editing because it is flexible enough to cover most requests. That split matches what we wrote about Qwen Image 2.1 as an edit model, where instruction editing, not first-frame generation, was the thing it did well.

The licence is the part that decides whether you can ship

P-H3-Image is a finetune of MiniMax H3 and carries the MiniMax H3 Community License. Its grant runs solely within the applicable territory, which excludes the European Union, the United Kingdom, the Republic of Korea and the United States unless you hold a separate licence from MiniMax. Commercial products and services must prominently display "MiniMax H3" on their user interface, and past 20 million US dollars in yearly revenue you need separate written authorization. The "Powered by MiniMax H3" notice sits in the encouraged list, next to adding an AI-generation identifier and writing a public post about your experience, so treat it as good practice rather than a contractual obligation. Public sharing is its own category: the Acceptable Use Policy (Exhibit A, item 12) prohibits disseminating machine-generated content in any public environment without clearly and prominently disclosing that it is machine-generated, so that disclosure is an obligation whenever you publish outputs, while the file identifier itself stays in the encouraged column. The text encoder's Qwen3-VL-4B parts and the Qwen3.5-9B upsampler are Apache 2.0, which is a different licence covering different files.

Read that as three separate questions. Territory comes first, and it applies to every use rather than only to products: the grant covers the applicable territory alone and the acceptable-use policy lists use outside it as prohibited, so a personal experiment in New York or Berlin sits outside the licence instead of being a small-scale version of it. Client work adds delivery on top of that, so if you or the client are in one of those four regions, it is a conversation with MiniMax before the invoice, not after. Shipping a product adds the UI notice and the revenue threshold, and being a finetune means the restriction travels with the derivative instead of washing out. I am describing terms, not giving legal advice, and for commercial use the licence file itself is the thing to read, preferably with a lawyer.

What I would do with it

  • Decide on the packaging before anything else. A 24 GB card means the w4a8 set with an int8 VAE, which is the only one of the three that fits; the int8 set peaks near 31 GB, so it wants 48 GB. A bf16 set is either a 56 GB Comfy job or a 76 GB offloaded diffusers job, which is a rented datacenter card.
  • Keep the upsampler on, and save the prompt it returns next to every render, with the seed and canvas size. That set is what makes a one-line request reproducible.
  • Write the keep list on the first attempt. Iterating without one is how you end up with three variations of a different picture.
  • Inspect at 100% on a face, a hand and the background. The point of 8 MP is the crop, and a full-frame thumbnail will not show you the failure.
  • Settle the licence question before you promise a client anything.

None of this makes the model a replacement for a hosted editor, and it is not trying to be. It is an open-weight editor that renders studio-resolution stills and asks for a serious card and a licence review in return. If you have the card and the territory works, it is the most interesting image release of the past two weeks. If you do not, a hosted editor is the other side of that trade, and it exists because "free" always has a bill attached to it somewhere.

Sources

Share this guide
X LinkedIn