MiniMax H3 Image vs P-H3-Image

Two image models shipped a week apart, and the names make them look like a fast mode and a full mode of the same checkpoint. They are different pieces of work.

MiniMax-H3-Image is MiniMax's own image model. The license is dated November 10, 2026, and the research post is dated November 11. It samples in six steps, aims at about 16 megapixels, and accepts up to 16 reference images. P-H3-Image is Pruna's still-image model, published November 3, built by reworking the MiniMax-H3 video model so a single frame comes out sharp. It samples in eight steps, and the current model card aims it at 8 megapixels with up to eight input images.

Use MiniMax-H3-Image when the file you keep is a large still, or an edit that has to respect a long list of reference photos, and a hosted generator can hold the GPU. Use P-H3-Image when you want the weights on your own machine and a 4-bit ComfyUI setup at about 21 GB of VRAM is the computer you actually have.

One chart has already started circulating as a five-times speedup. Read the legend before you apply it here. The slow bars are MiniMax-H3, the video model, run as a still. They are not MiniMax-H3-Image.

What the two names refer to

MiniMax H3, announced July 31, 2026, is a video model. It reads text, images, video, and audio, and it writes moving clips. Pruna started from those weights because a video model has already had to learn materials, light, and bodies that stay coherent from frame to frame. Left alone, that checkpoint can emit a single image, and Pruna's account of the result is blunt: soft, blurry, and too large for most people to run.

On November 3 they published P-H3-Image. The rework covers the parts a still actually uses. A prompt upsampler, based on Qwen3.5-9B, turns a short request into the kind of prompt the model renders well. The text encoder is distilled from the original 32B encoder down to 4B. The diffusion schedule, tuned for video, is adapted so fine detail survives. Sampling drops from the video model's 50 steps to 8, with no classifier-free guidance. A new decoder turns one frame into one image, instead of decoding video a few frames at a time.

A week later MiniMax published a different artifact. The MiniMax-H3-Image model card describes a 33.1B diffusion transformer with the H3 audio branch removed, the original MiniMax-H3 text encoder left unchanged, and distillation to six steps with no classifier-free guidance. Text-to-image and editing share that one checkpoint. The text encoder is not in the image repository. Loading the model pulls it from the MiniMax-H3 weights, about 66.7 GB, unless you already have that folder locally.

The family tree is short. Video model, then Pruna's image fork of that video model, then MiniMax's own image checkpoint. P-H3-Image shipped first. It is not a compressed copy of MiniMax-H3-Image.

The numbers, with the disagreements left visible

Two public descriptions of P-H3-Image do not match, and a buying decision should not average them.

The P-H3-Image model card lists a 33.1B transformer and about 100 GB of bf16 weights. Pruna's November 3 post describes the backbone as 20B after replacing the block that reads the timestep, and a 55 GB bf16 download against 56 GB of memory at 8 MP. The post says "up to 9" input images in the opening and "up to 8" further down. The card says 8. For the diffusers pipeline you clone today, follow the card. For the quantized files, both pages agree, and the ComfyUI weights are the set that peaks near 21 GB. Recheck both pages if you are ordering a GPU from this table. The figures move.

Speed claims sit on different baselines

Gray is MiniMax-H3. Purple is P-H3-Image. On one B200 in bf16, Pruna reports 12.4 seconds against 2.6 seconds at 4 MP, and 30.6 seconds against 5.8 seconds at 8 MP. The third row is a different test: on an L40S, the video model is bf16 and offloaded, while P-H3-Image is 4-bit, 19 seconds against 6.3 seconds at 4 MP. Pruna says the B200 rows compare 50 steps with 8, and that those runs were not further compiled.

That is a real improvement if your alternative is the video checkpoint used as a camera. It says nothing about MiniMax-H3-Image, which was distilled to six steps on its own and was not in that test.

MiniMax's launch post says a 4096×4096 H3-Image frame comes back in under 10 seconds, and that a Highspeed mode is faster than the standard pass. The post does not name the GPU, the batch size, or whether prompt expansion is inside the clock. Separately, the P-H3-Image card says about 9 seconds for an 8 MP frame on one B200 when the prompt is used as written, and another 6 to 11 seconds when the upsampler runs. Those sentences use different models, different canvases, and different prompt paths. Subtracting them produces a number nobody measured.

If latency decides a purchase, time the path you will call: your resolution, your prompt length, and the machine that will actually serve it.

Resolution is the quality setting

MiniMax-H3-Image's useful range is about 4 to 20 megapixels, and the card says results are generally best around 16. A square frame at that area is 4096×4096. A 16:9 frame is 5472×3072. The image generation guide treats 4K as an area of 4096×4096 pixels, which is why a 16:9 "4K" file is 5472×3072 and not 3840×2160. You can pass an exact width and height instead. The area has to stay between 655,360 and 20,000,000 pixels, and the aspect ratio between 1:3 and 3:1. The server rejects a size outside that box. It does not quietly scale it.

The sampler is just as strict. Omit the step count or set it to 6. Omit guidance or set it to 1. Anything else returns an error. Negative prompts are rejected. A smaller canvas can be drawn on a larger one and resized back, but the sizes in the 16 MP table already sit above that floor. Generate the size you mean to keep.

P-H3-Image's card tells you to judge text-to-image at 8 MP. The default square is 2912×2912, and 16:9 is 3872×2176. Edits are aimed at 4 or 8 MP. Sides are rounded to a multiple of 32. The card is plain about the small sizes: at 2 and 4 MP, faces, hands, and small details break more often.

The ladder is that warning in one picture. At the small end the face goes smooth and waxy. From 8 MP upward, skin, stray hairs, and small jewelry start to hold. The framing can also move between sizes, because the noise field changes with the canvas. A composition you liked at preview size is a new roll of the dice at 8 MP, which is the size worth judging.

What each model is showing off

These are the examples each maker published. They are selected to present the model well. I did not run a paired shootout, so the right next step with your own prompt is still a paired shootout.

MiniMax's photographs lean on ordinary light: a moon behind a sheet, a face in a dark room, birds in a wet flat. The launch note describes the target as texture that stays specific, closer to a real scene than to a smoothed render. The same showcase includes a harder claim, logic puzzles in Chinese and English solved correctly on the board. That is their evidence for text that has to be reasoned, not just letter-shaped. Check it on a sentence you actually need in the picture.

P-H3-Image's published strengths are documentary and editorial photographs, portraits with visible skin, film frames, TV-frame anime, ink and charcoal, screen prints, and flat graphics with short quoted text. The prompting rules are specific enough to copy.

Say what the image is, in the first sentence. A photograph by a street photographer, a frame from a Japanese anime series, a two-color risograph, a flat scan of a handwritten recipe card. That sentence sets the medium more reliably than a style adjective. Then describe one scene: a viewpoint, one subject caught in a moment, a few supporting objects, and light with a source and a direction.

Spell every word that must appear, in quotes, and say where it sits. End with "no other text." If you want no text, do not mention signs or labels at all. Leave out stunning, epic, cinematic, vibrant, hyper-detailed, 8k, and masterpiece. Those words push this model toward a glossy, oversaturated, airbrushed picture. A very short prompt is weak unless the upsampler rewrites it. A very long prompt, or a prompt full of praise, gets worse.

Hands can still break. Small people in the background can distort. Short quoted strings are the text it handles. Body copy and code turn into pseudo-text, and longer passages sometimes miss a letter.

Editing is a different instruction

Both models edit from a written instruction, and both will redraw a scene when the instruction is vague. Capacity and prompt shape are where they split.

MiniMax-H3-Image takes up to 16 reference images. Upload order is how the prompt numbers them. One published edit starts from an empty room plus nine pieces of furniture and assembles a living room.

On the API, a short prompt in any language can be expanded before generation. Locally, the same job belongs to the optional prompt-expander weights, or to a prompt you write out yourself. The model expects a detailed English description of the whole image.

P-H3-Image takes a list of up to eight. The first image is the one being changed. The others are references the prompt can name as Picture 2, Picture 3, and so on. A line such as "make the pickup truck red" often causes a full redraw. The upsampler exists to turn that line into an edit brief that names what stays.

With the upsampler removed, Pruna's template is the one to follow. "Edit Picture 1. The pickup truck is red. Keep everything else exactly as it is." Point at the object and where it is. The faded blue cap, not "his hat." Describe the finished look, and give a new object a size next to something already in the frame. List the faces, the pose, the background, and the light that have to survive. New text goes in quotes, next to the words that must stay.

The same path can cast a scene. Eight portraits and one sentence become a dinner party, after the upsampler says what to take from each picture.

Eight references already cover a small cast. Sixteen matters when each object has its own photo: a room of specific furniture, a product layout, a wardrobe board where every garment is a real picture. If your edit is one photo and a color change, either model can take the job, and the prompt that says what must stay will matter more than the unused headroom.

What you need in the room

Running MiniMax-H3-Image yourself means the image weights, the separate text encoder, and a serving stack. The card points at SGLang, vLLM-Omni, diffusers, and ComfyUI. Its hardware examples are an H200, an RTX PRO 6000 Blackwell, and an AMD MI350X, with multi-GPU sequence parallelism when you want the latency down. That is a server plan.

The P-H3-Image diffusers pipeline in bf16 is the same class of plan. The card asks for about 101 GB with the upsampler at 8 MP, or about 83 GB without it, and says an 80 GB GPU can run with CPU offload at a peak around 76 GB. The setup that fits a smaller machine is the 4-bit ComfyUI set: about 21 GB of VRAM at 8 MP, and ComfyUI loads it without custom nodes. Keep the canvas at 8 MP. Pruna says the 4-bit weights can move small details and sometimes the whole composition, while the 8-bit set stays closer to bf16.

When the official checkpoint is the one you want, and that server list is not your setup, call a hosted generator. MiniMax H3 Text to Image runs the model from a prompt at the 1K, 2K, and 4K area presets, with a Highspeed variant for the passes where the wording is still moving. When the job is a change to a picture you already have, or a scene built from several references, MiniMax H3 Image to Image takes the instruction and up to 16 images. Use the 4K area when the file has to hold detail. PNG, JPEG, and WebP are the output formats.

The license is part of the choice

Both weight releases draw the same map. The MiniMax-H3-Image community license, dated November 10, 2026, sets the applicable territory as worldwide except the European Union, the United Kingdom, the Republic of Korea, and the United States. Using the works through a hosted service sits inside that agreement. MiniMax points to a separate written authorization for use beyond the grant.

P-H3-Image is released under the MiniMax H3 community license. Pruna's card summarizes the same four excluded territories, and it says products built on the model have to display "Powered by MiniMax H3." The 4B text-encoder parts and the Qwen3.5-9B upsampler are under Apache 2.0. The ComfyUI nodes are MIT.

This is not a legal opinion. If you are in one of those four places, or you are about to ship a product on either set of weights, read the license that matches the files you will load. MiniMax's own API and the Hailuo app are their hosted products, on their terms, which is a different document from the community license.

Which one to open

Open MiniMax-H3-Image for a finished still around 16 MP, for an edit that has to carry a long list of reference photos, and for a workflow where someone else holds the GPU. Set the area to 4K. Write the scene as a place you could walk into: subject, viewpoint, light with a direction, and the exact words in quotes.

Open P-H3-Image for local iteration on a smaller GPU. Use the 4-bit ComfyUI weights, leave the upsampler on, and judge the picture at 8 MP. It fits a film still, an editorial portrait, or a graphic with a short line of quoted text, within the limits the card already prints: hands, dense text, and anything rendered too small.

Skip both downloads if the territory clause blocks you and you do not have written authorization. The decision then moves to MiniMax's own API terms.

The H3 in both names is the architecture family. It does not carry the step count, the megapixel target, or the speed chart from one checkpoint to the other.

Questions people bring to this comparison

Is P-H3-Image a faster MiniMax-H3-Image?

P-H3-Image comes from the MiniMax-H3 video model and was published on November 3, 2026. MiniMax-H3-Image is MiniMax's own image checkpoint, with a November 10 license, six sampling steps, and a target around 16 megapixels.

Where does the "five times faster" figure come from?

Pruna's B200 test compares P-H3-Image at eight steps with MiniMax-H3 stills at fifty steps. At 8 MP the times are 5.8 seconds and 30.6 seconds. MiniMax-H3-Image was not in that test.

Should a small preview be upscaled to 4K?

Sample the size you intend to keep. Pruna's resolution ladder shows detail, and sometimes the composition, changing as the canvas grows. MiniMax-H3-Image is built to generate near 16 MP directly, and its API rejects an out-of-range size instead of rescaling it for you.