MiniMax H3 Image: what we know

First, the name trap. "MiniMax H3" is the video model (also sold as Hailuo 3): 5 to 15 second clips with sound [reported]. H3 Image is the still-image member of the same H3 family, built on the same architecture minus the audio layers [reported]. MiniMax announced it on November 10, 2026, in a blog post titled "Bringing 4K Image Generation into the 10-Second Era" [official], and the weights went public on Hugging Face the same day [reported]. Search results for "minimax h3 image" mix video-model pages, free-video guides, and image-model pages freely — and Google autocomplete already returns 10 real suggestions for the phrase, led by comfyui, gguf, prompting guide, 4k speed, reddit, and edit [reported]. If a page quotes per-second prices, clip lengths, or a 768p cap, it is about the video model, not this one.

The headline spec is resolution without an upscaler. H3 Image generates natively at around 16 megapixels at every aspect ratio — 4096×4096 at 1:1, 5472×3072 at 16:9, 6272×2688 at 21:9 — with the API capping output at 20 megapixels [official]. The model is a 33.1-billion-parameter dense diffusion transformer, distilled to six sampling steps with no classifier-free guidance [reported]. The second spec is control: text-to-image, reference-based generation, and editing in one checkpoint, with up to 16 reference images addressed positionally as <Picture 1> through <Picture 16> in upload order [reported]. Prompts can run to 32,000 characters through the API, so a full poster brief goes in as written [official].

The third spec is text. The model renders long, dense text for posters, infographics, and UI screens [official], and early community feedback singles out instruction following — the model tends to render what the prompt actually asks for, including layout [reported]. That is a vendor-adjacent claim so far: MiniMax's own materials carry no benchmark score and no leaderboard entry for the model exists yet [reported]. The one independent test in circulation is Pixel Dojo running the same eight prompts on H3 Image and GPT Image 2.6, one try each — H3 Image made the more natural photo, GPT Image 2.6 was better at text and layout [reported]. Eight prompts, one try each, is a sniff test, not a verdict.

The speed claim needs a footnote of its own. The launch post says a 4096×4096 image lands in under 10 seconds [official], while the announcement post on X words it as 10 to 20 seconds [reported] — and MiniMax's own comparison chart shows 4K text-to-image at a 16.0 second median in Standard mode and 10.3 seconds even in Highspeed mode [reported]. Community local runs are slower still, around 56 to 70 seconds on an RTX 3090 in early tests [reported]. Treat "under 10 seconds" as MiniMax's serving-stack claim, not a number your setup will reproduce.

The fastest free route is the fal playground. fal hosts H3 Image with a generator on the model page, and free runs depend on the active offer and your eligibility — the page shows your remaining allowance and any refill time before you generate, and when free runs are unavailable you continue with fal credits at the model's usual price [official]. Do not import the video model's numbers here: the widely quoted "5 free generations a day" sandbox is fal's H3 Max video offer, not H3 Image [reported]. There is no fixed daily free count for the image model that we could confirm [unknown].

The second free route is local. The weights are public on Hugging Face as MiniMaxAI/MiniMax-H3-Image, so running the model costs only hardware and electricity [official]. ComfyUI shipped day-0 native support with ready-made workflows for the diffusion model, text encoder, and VAE components [reported]. The download is not small: the image model is 68.65 GB, and it needs the text encoder from the H3 video model's repository, another 66.7 GB [reported]. And the license is the catch — see the license section below before you download for anything commercial.

Hailuo AI, MiniMax's consumer app, lists H3 Image as available [official], but no free-credit terms for the image model have been confirmed — the welcome packs and trial offers quoted in "H3 free" guides are for the video model [unknown]. A daily free-credit offer on gen-image could not be verified as of November 12, 2026 [unknown]. The practical free test: generate one 4K poster with a headline and one photo edit from two references on fal, then check the text spelling and whether the untouched areas stayed put. That is the claim that matters, and it costs nothing to check.

The official API is live on MiniMax's API Platform under model IDs MiniMax-H3-Image and MiniMax-H3-Image-Highspeed, with text-to-image and image-edit endpoints [reported]. Billing is by token, not by image: $2.30 per million input text tokens, $1.50 per million input image tokens, and $3.80 per million output image tokens on the standard model; Highspeed runs $2.30, $3.00, and $7.60 for the same three [reported]. The API accepts any size up to 20 million pixels, 1 to 10 images per request, and JPEG, PNG, or WebP output [reported].

fal mirrors the same model as minimax/h3-image text-to-image and image-edit endpoints, each in a standard and a Highspeed version, pay-per-use through the API [official]. Three limits from the API guide are worth knowing before you build: preview images, masked edits, and transparent backgrounds are not supported [reported]; image edits always apply prompt enhancement, which cannot be turned off [reported]; and neither endpoint is idempotent, so a request that times out on your side is still charged if the server finishes it — MiniMax tells developers to turn off automatic retries [reported]. The server waits up to 600 seconds, well above the 10 to 20 seconds in the announcement [reported].

MiniMax publishes example costs per image, with a caution attached: the examples "do not represent fixed per-image prices or average costs" and "actual charges are based on Token usage" [reported]. On the standard model, the examples are about $0.02179 for a 1K image, $0.02333 for 2K, and $0.06879 for 4K; Highspeed lists $0.03529, $0.03622, and $0.13001 for the same sizes [reported]. A 4K edit from one 1K reference image is listed at about $0.07312, and from three references at about $0.07914 [reported]. The 4K image holds 16 times the pixels of the 1K image and costs about 3.2 times as much at list price [reported].

The launch-week discount is real but time-boxed. MiniMax's blog offers 50% off the standard price for image generation at 4K resolution and above during launch week [official], and the price list sets the end at November 17, 2026, 16:00 UTC, for requests delivering at least 16,600,000 pixels — a 4096×4096 image qualifies, a 2K image does not [reported]. In that window a 4K image costs about $0.03439 at the example rate [reported]. MiniMax also frames the general price as 50–70% of comparable mainstream offerings [official], without naming the baseline, so that claim cannot answer which model is cheaper. After November 17, budget the list example prices until MiniMax says otherwise.

Third-party hosts set their own rates on top. kie.ai lists H3 Image at $0.0275 per 4K image and $0.009 per 1K/2K image against the official examples of $0.06879 and $0.02179 [reported]. Venice charges $0.04 per image at 1K and 2K and $0.10 at 4K [reported]. Those are host prices for host endpoints — convenient for prototyping, not a vendor list price.

The weights are open, but they are not under a standard permissive license. H3 Image ships under the MiniMax-H3-Image Community License Agreement, dated November 10, 2026 [reported]. The agreement's applicable territory excludes the European Union, the United Kingdom, the Republic of Korea, and the United States of America — people there are told to contact MiniMax about a separate license [reported]. That exclusion is the single most reported fact about this release after the 4K claim, and it is why hosted access matters: running the model through a hosted API sidesteps the self-hosting restriction [reported].

The commercial limits go further. Written authorization from MiniMax is required if the annualized revenue of you or your affiliates exceeds $1 million, and the same applies to anyone who offers the model to others as a hosted service [reported]. A commercial product that uses the model must display the name MiniMax-H3-Image in its interface, and the license bars using the model's works or outputs to improve any other AI model [reported]. On the other side, MiniMax claims no rights over the outputs you generate [reported]. None of this is legal advice — the license text on the Hugging Face repository is the authority, and it is the document to read before any commercial use.

The two November image launches solve different problems. Seedream 5.5, ByteDance's model from November 9, 2026, is built for editing depth: up to 100 rounds of continuous edits on one image, with unedited areas preserved at the pixel level, and native RGBA transparent output [reported]. H3 Image is built for resolution and openness: native 16MP 4K straight out of the model, up to 16 reference images in one edit, and open weights you can run locally [reported]. H3 Image explicitly does not support transparent backgrounds or masked edits [reported] — the two things Seedream 5.5 is best at.

On price, H3 Image is the more transparent of the two: MiniMax publishes example costs of about $0.02179 per 1K image and $0.06879 per 4K image [reported], while ByteDance has published no USD list price for Seedream 5.5 [reported]. On editing, Seedream 5.5 is the deeper tool; on stills at full resolution, H3 Image is the cheaper way to a 4K file. For video creators, the practical split is simple: use H3 Image (or its ComfyUI workflow) to make high-resolution keyframes and reference stills, and Seedream 5.5 to iterate a poster or product shot through many rounds. Against GPT Image 2.6, the only direct test so far is Pixel Dojo's eight-prompt run: H3 Image made the more natural photo, GPT Image 2.6 was better at text and layout [reported] — a sniff test, not a verdict.

Four things, in order of importance. First, November 17, 2026, 16:00 UTC, when the 50% launch discount on 4K-and-above generation ends — the price after that date is the list example price until MiniMax says otherwise [reported]. Second, independent verification of the speed and quality claims: MiniMax's own numbers already disagree with each other (under 10 seconds in the blog, 10 to 20 seconds in the announcement post, 16.0 seconds Standard median on its own chart) [reported], and no independent benchmark or leaderboard entry exists yet [reported]. Third, licenses for the excluded territories — whether MiniMax grants them, and on what terms [reported]. Fourth, the ComfyUI ecosystem: local workflows, GGUF quants, and prompting guides are already the top autocomplete suggestions, and the quality of that community tooling will decide whether the open weights matter beyond the launch week [reported].

One positioning note, stated plainly: this is an image-model guide on a video site. It is here for the same reason as our Seedream 5.5 guide — video creators make keyframes, reference stills, posters, and thumbnails before they animate anything, and a 16MP still with 16 reference images is a video workflow input. The video comparisons remain the site's core; this page is a tool for the step before the video.