MiniMax H3-Image generates and edits still images. This is the standard H3-Image version released on 10 November 2026, alongside H3-Image Open Weights and H3-Image Open Weights LoRA. In our NSFW test on 11 November, all 36 requests returned images without refusals. Mild, Moderate, Strong and Maximum were delivered as asked; Very strong was toned down in two of three text-to-image runs and all three edits.
Editorial portrait of a man in his forties with a salt-and-pepper beard and a navy wool turtleneck, sitting by a large window on a grey morning, soft natural light, medium format film, natural skin texture.
A jazz singer in a midnight-blue sequined gown on a small club stage, warm amber spotlight, a brass band blurred behind her, smoke in the air, shot on 35mm film, cinematic wide frame.
A minimalist perfume bottle on a black marble plinth, the glass label reads "NOIR 07" in thin gold serif letters, single soft key light, luxury product photography.
Furnish the empty room in image 1 with a mustard velvet sofa, a round walnut coffee table and a large green plant, and put the dog from image 2 lying on a wool rug in front of the sofa. Keep the window light.
Move the chef in image 1 to a night-market street food stall in the rain, cooking noodles in a wok under string lights. Keep his face, beard and white jacket.
Pick it for 2k pictures at the 1k price, 4k output at $0.108, or edits combining up to 16 inputs. Skip it when you need to reproduce a generation with a seed: choose Open Weights instead, which also costs less at 1k. Choose Open Weights LoRA when you want to load your own style files. None of the three delivered Very strong edits as asked in our three-run samples.
How it scores. The score card draws its values from our recorded tests; the observations below explain what those records mean for choosing a version.
Price: text to image costs $0.036 at 1k or 2k and $0.108 at 4k. Edits use the same output price plus $0.0096 per input image, so a 1k edit with two inputs costs $0.0552. Failed and refused requests are not charged.
Generation parameters:prompt up to 5,000 characters, aspect_ratio with 15 options, resolution 1k/2k/4k with 1k as the default, and output_format jpeg/png/webp. Each request returns one image.
Edit parameters: generation settings plus image_urls with 1 to 16 inputs. Refer to them as “image 1”, “image 2” and so on in the prompt.
Sizes: measured outputs included 1024 × 1024 at 1k square, 2720 × 1536 at 2k 16:9, 4096 × 4096 at 4k square and 5472 × 3072 at 4k 16:9. Ratios: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9, 9:21, 2:1, 1:2, 3:1, 1:3, 5:4 and 4:5.
Not available: seed, LoRA input, negative prompt, mask or transparent background.
Speed: reported model time was 12 to 20 seconds at 1k across two sessions on 11 November, and 22 seconds for one 4k image. Queue and download time are additional. Speed changed between sessions on the same day, particularly for Open Weights, so these timings are observations rather than fixed waits.
良かった点と足りなかった点
良かった点
2k keeps the 1k price. Both cost $0.036 per generated image, compared with $0.048 for Open Weights at 2k.
4k costs $0.108. Open Weights charges $0.144 for a generated 4k image.
Edits accept up to 16 images. That leaves room to combine a setting, a subject and further references in one request.
Short text came out right. The same-prompt wooden sign reproduced “HOT SOUP 3 FRANCS” correctly.
Viewpoint and labels followed more of the request. H3-Image showed the elevated street view and included more botanical labels than the other two models in the comparison, although the species name was misspelled.
Every content-test request returned an image. There were no refusals across 36 runs on the two endpoints.
足りなかった点
No seed. You cannot hold a generated picture steady through a seed while changing one part of its prompt.
Very strong often came back toned down. Two of three text-to-image runs and all three edits did not show the requested level in full.
1k costs more than Open Weights. Generation is $0.036 here against $0.024 there.
Several editing controls are absent. There is no negative prompt, mask or transparent-background setting.
One image per request. A request cannot return a batch of variations.
What the images look like
The gallery uses six published examples from the model page, all suitable for public display. Content-test images are excluded; their outcomes appear in words and in the table below.
4k portrait. The bearded-turtleneck-4k example is a window-lit portrait of a man in a navy wool turtleneck. It illustrates the 4k generation setting; the gallery copy is resized for display.
Stage and label. Midnight-jazz-sign presents a singer on a small club stage under amber light. Perfume-label-noir puts a glass bottle on a dark plinth with the short “NOIR 07” label. The perfume example concerns text in the frame; the jazz prompt below specifies the stage scene without requesting lettering.
Two-image room edit. Dog-furnished-room combines an empty-room input and a dog input. The edit adds furniture and places the dog on a rug while retaining the room's window light.
Subject-preserving edits. Chef-night-market changes the chef's setting to a rainy food stall while keeping his face, beard and white jacket. Sneaker-terracotta changes the shoe's colours while keeping its shape and the concrete floor.
Same-prompt comparison. We also compared six prompts at 1k on 11 November, using seed 7 on the open builds. H3-Image, Open Weights and P-H3-Image all spelled “HOT SOUP 3 FRANCS” correctly. H3-Image included more of the botanical labels, but all three misspelled the species name. It also followed the requested view from a pedestrian bridge, where the other two showed ground-level or under-bridge views. Open Weights turned a requested coffee dripper into a mug with a handle. These are observations from those prompts, rather than a guarantee for other scenes.
Prompts that worked
Quote any wording that must appear in the picture and describe where it belongs. For edits, call the sources “image 1” and “image 2”, then say what should stay as well as what should change. The examples below preserve the prompts used for the published pictures.
Perfume label. Text to image, aspect_ratio 1:1, resolution 1k.
text
A minimalist perfume bottle on a black marble plinth, the glass label reads "NOIR 07" in thin gold serif letters, single soft key light, luxury product photography.
Jazz stage. Text to image, aspect_ratio 16:9, resolution 2k.
text
A jazz singer in a midnight-blue sequined gown on a small club stage, warm amber spotlight, a brass band blurred behind her, smoke in the air, shot on 35mm film, cinematic wide frame.
Furnish the empty room in image 1 with a mustard velvet sofa, a round walnut coffee table and a large green plant, and put the dog from image 2 lying on a wool rug in front of the sofa. Keep the window light.
SpicyAPI adds no content filter. On 11 November 2026, we ran 36 requests across text to image and edit, covering five levels from Mild to Maximum. The tests used adult subjects only, and every returned image was checked by eye. We publish the level names and outcomes, but keep the test prompts unpublished. Each endpoint had three runs per level except Maximum, which used two probe types for six runs.
Level
Text to image
Edit
Mild
3 of 3
3 of 3
Moderate
3 of 3
3 of 3
Strong
3 of 3
3 of 3
Very strong
1 of 3, 2 toned down
0 of 3, 3 toned down
Maximum
6 of 6
6 of 6
Reading the table. Counts show images delivered as asked, rather than requests accepted: all 36 returned images and were billed. Mild, Moderate, Strong and Maximum matched the requested level in every run, while Very strong was toned down in most text-to-image runs and every edit. Passing Maximum therefore did not mean that every level below it was delivered consistently.
Additional Very strong text-to-image checks brought the two-person total to 1 of 6 delivered as asked; single-person requests delivered 2 of 3. These small samples record one day, not a promise for every prompt. If the top levels matter for your edits, consider the alternatives below and their recorded limitations; switching to an open-weight sibling did not solve the Very strong edit result in this test.
Adults only in every prompt and every source photo; no real person's likeness without consent; nobody who could read as a minor. You are responsible for what you generate and publish. Generated media is kept for 14 days and then deleted.
Alternatives
MiniMax H3-Image Open Weights: choose it for a seed, 1.5k and 3k settings, or $0.024 generation at 1k. It takes up to nine edit inputs. At 2k and 4k, generation costs $0.048 and $0.144, compared with $0.036 and $0.108 here. Its three Very strong edits were also toned down.
P-H3-Image: a different model with 1k to 4k output. One 4k image took 30 seconds in our run. Matching the prompt, seed and ratio on P-H3-Image and Open Weights produced different pictures.
Seedream 5.5: starts at $0.0195 for 1k, supports up to 4k, ten edit references, transparent PNG and layer splitting. Its 7 November test recorded no refusals, with two edits toned down. The Seedream 5.5 review gives the level-by-level results.
Seedream 4.5: generation costs $0.04 at either 2k or 4k, and edits take up to 14 references. See the Seedream 4.5 review for its tested behaviour.
SpicyAPI adds no content filter. All 36 requests in our 11 November test returned images, though Very strong was toned down in two of three text-to-image runs and all three edits. The consent, age and responsibility rules above apply to every request.
The current price schedule puts both 1k and 2k generation at $0.036; 4k costs $0.108. Equal pricing is the listed rate, rather than a claim that the pixel dimensions are equal. Edits add $0.0096 per input image to the output price. Failed and refused requests are not charged.
Choose 1k, 2k or 4k across 15 aspect ratios. Our square 4k output measured 4096 × 4096, or 16,777,216 pixels; 16:9 measured 5472 × 3072, or 16,809,984 pixels. The dimensions depend on the ratio. A gallery preview can be smaller than the generated file.
This version has no seed, accepts up to 16 edit inputs and offers 1k, 2k and 4k. Open Weights adds a seed and 1.5k/3k sizes, with up to nine edit inputs. Open Weights costs less at 1k; this version costs less at 2k and 4k. The content table does not establish that either is more permissive overall.
No. P-H3-Image is a different model. The older MiniMax H3 Image LoRA makes stills with the H3 video model at 1k or 2k; it is a separate family from H3-Image. Its review covers that older product. H3-Image Open Weights LoRA is the newer sibling for loading your own LoRA files.
We recorded 12 to 20 seconds of model time at 1k across two sessions on 11 November, and 22 seconds for one 4k image. Add queue and download time. Speed varied between sessions that day, so allow for a different wait on another run.
Pick Seedream 5.5 for exact lettering, transparent cutouts and layers from $0.0195; skip it if you need a seed, a mask or more than 10 edit references.