MiniMax H3 Image vs. GPT Image 2.6 hands-on comparison: cheaper and faster, but still behind on complex instructions
The author compares MiniMax H3 Image and GPT Image 2.6 across six rounds of hands-on testing. H3 Image has advantages in image realism, generation speed and price—a 4K 16MP image costs about 0.2261 yuan per image with the first-week discount—but it still falls short of GPT Image 2.6 in executing complex instructions, precise quantity control and commercial poster design. H3 Image is derived from the H3 multimodal model, supports text-to-image, image editing and up to 16 reference image inputs, and its model weights have also been released. The author also notes that MiniMax's image capabilities are scattered across different entry points such as Hailuo and MiniMax Code, so the experience for ordinary creators is less coherent than the ChatGPT ecosystem.
Why read this
AI score 65
The author compares the two image generation models across six rounds of hands-on testing, illustrating the trade-off between realism and instruction precision, and adds differences in pricing and product entry points.
Topics
Body · AI translation
Text | Fourth Wave, by Feng Haoqin
English text is being prepared.
Before the formal review, let's first get to know a few core feature points of H3 Image.
It is derived from the earlier H3 multimodal model. In addition to text-to-image and image editing, it also supports input of up to 16 reference images. The official emphasis is on several advantages: generation speed, image realism, complex instruction-following capability, and low price. A roughly 16-megapixel 4K image takes as little as 10 seconds, at a price of less than 0.3 yuan.
But a few days earlier, on November 4, OpenAI served up GPT Image 2.6, which also focused its upgrades on generation speed, image editing, and visual consistency. With the two models released back to back, it is hard not to wonder: has MiniMax actually caught up with GPT this time?
To find out, Fourth Wave ran several rounds of practical tests with the two models. From slice-of-life documentary and tech covers to Chinese posters, industrial exploded views, multi-reference image fusion, and even AAA game scenes, all were done at 2K resolution, to see just how far apart they really are.
Image realism: MiniMax really has something here
In the first round, we asked the two models to generate a post-rain street scene in an old residential area of Shanghai. The image needed an old man repairing a bicycle, a young woman passing by carrying a plastic bag, and pedestrians talking in the distance. No cinematic filter, and no deliberately staged dramatic lighting.
(Left/GPT Image 2.6 Right/MiniMax H3 Image)
GPT Image 2.6 performed quite steadily. The old man, the young woman, and the distant pedestrians basically all appeared in the specified positions, and environmental elements such as the buildings, convenience store, and electric scooters were also fairly complete.
But the whole image was too neat and tidy. The street, the people, and the light were all just right, which instead took away some of the casualness of real life.
By comparison, the MiniMax H3 Image picture had more of the rough texture of real life.
The wrinkles on the old man's face, his rough hands, the rust on the bicycle, and the mottled walls all had a strong sense of texture. The overall image was rougher and had more of the flavor of street humanist documentary photography.
But it also showed obvious deviations. The request was for a young woman, but the generated figure clearly looked too mature, and the specified convenience store was not clearly presented either.
H3 Image can indeed already generate documentary photos with a considerable sense of life, but while pursuing realistic texture, it may inevitably sacrifice accurate execution of specific instructions. This may also be one reason why quite a few developers, after using it, commented that it "hardly has any AI flavor."
So in this round, H3 Image's realistic texture is worth affirming, but as a review, GPT Image 2.6 executed the instructions more accurately and, on balance, held a slight lead.
(Left/GPT Image 2.6 Right/MiniMax H3 Image)
GPT Image 2.6 placed data centers, transmission towers, and substation facilities in the same scene, with complete structure and a clear theme, but the overall result looked more like a polished commercial rendering, and its design language was not particularly novel.
H3 Image, by contrast, had more of its own style.
It chose a large area of blank sky, with buildings and transmission lines concentrated in the lower part of the image. The gray-brown tone was restrained and natural, making the whole thing closer to documentary photography and giving it more of the visual atmosphere of a magazine cover.
In this round, if the emphasis is on information expression, GPT Image is more accurate, but if looking at aesthetics and photographic atmosphere, I lean toward H3 Image.
From the first two rounds, it can be seen that MiniMax H3 Image indeed has the ability to go toe to toe with leading image generation models. The model no longer relies only on low prices to attract users; its image texture and aesthetics are genuinely competitive.
When actually put to work, the gap shows
Next, we began to raise the difficulty.
In the third round, we started using the two models to design Chinese commercial posters. The requirements included the complete conference name, time, location, and six conference agenda items, with no typos or missing information.
The result was somewhat unexpected: both models performed well in Chinese, and the specified text and time were all presented accurately.
But the gap in design... is hard to describe in a word.
(Left/GPT Image 2.6 Right/MiniMax H3 Image)
GPT Image 2.6 used a black-and-orange color scheme, with clear visual hierarchy for the title, agenda, and organizer information, and overall it already had the basic usability of a commercial poster. Although the top title was slightly crowded and the large orange graphic on the right was a bit overdone, at least a complete design idea was visible.
H3 Image went to the completely opposite extreme.
It laid out all the information neatly on an off-white background. The text was not a problem, and there was enough white space, but the visual expression was too simple. It was like a meeting notice, or an unfinished PowerPoint.
It understood the text requirements, but did not truly complete the design task. In this round, the outcome is presumably very obvious.
In the fourth round, we tested a more complex industrial product exploded view.
This time the requirements were very clear: a handheld searchlight, strictly split into 7 main components in order, with Chinese names accurately labeled, plus an additional display of 3 fixing screws and the complete assembled effect.
(Left/GPT Image 2.6, Right/MiniMax H3 Image)
GPT Image 2.6 completed almost all of these requirements. The 7 components were in the correct order, none of the 3 screws was missing, and the material distinctions among metal, glass, and circuit boards were also very clear. Although it cannot be used directly as an engineering drawing, it already has a decent professional presentation effect.
H3 Image's performance was not as ideal.
It also drew the 7 main components correctly, and the Chinese labels had no obvious errors, but it was asked for 3 screws and in the end gave only 2.
More noticeably, the finished product in the lower right corner also showed a large area of exposed silver heat-dissipation structure, which did not fully correspond to the black rear shell in the exploded view.
This kind of problem may be harmless in ordinary illustrations, but if it is used for product introductions or structural explanations, it cannot be ignored.
In the fifth round, we provided 4 different reference images, corresponding to a headphone product, a human subject, an action pose, and an architectural environment, and asked the models to fuse them into a high-end commercial advertisement.
(The reference images were the same for both models)
This time both models performed well.
(Left/GPT Image 2.6, Right/MiniMax H3 Image)
GPT Image 2.6's overall composition was more complete, and the person, headphones, and architectural environment were combined relatively naturally, with the texture of a mature commercial advertisement.
H3 Image also performed well in the person's skin, clothing texture, and lighting, especially the side profile and warm light, which had quite a fashion-photography feel.
But it did not fully preserve the original chair, and instead had the person sit directly on the edge of the building, so the scene arrangement deviated. As for the accuracy of restoring the person's identity and the appearance of the headphones, further judgment is needed in combination with the original reference materials.
Overall, in this round the two models each had their own strengths.
In addition, in the final round, we also added a set of 3A game scene tests, requiring the generation of third-person in-game screenshots in an abandoned industrial city.
(Left/GPT Image 2.6, Right/MiniMax H3 Image)
In the images generated by both models, the robots, damaged buildings, rainwater puddles after rain, and game UI were all quite refined, but overall they were still closer to carefully staged game promotional CG, lacking the camera language that naturally forms during actual gameplay.
By comparison, the interactive HTML game generated with ClaudeFable 5.2 was actually more immersive as a real game in the combination of weapon perspective, mission information, and scene layout.
Since we have mentioned this, we might as well also note that later we will use different Agent models to vibe such interactive 3A game works, so that everyone can see the characteristics of different models more intuitively from game videos. Interested friends can follow our subsequent videos.
This comparison at least shows one thing: without Agent-assisted scene construction, a single AI-generated image, even if its materials and lighting are refined enough, may not truly reproduce the in-game texture of a game.
So overall, after testing, the results are fairly clear.
H3 Image is competitive in realism and visual aesthetics, but when facing complex instructions, precise quantity control, and commercial design, GPT Image 2.6 is still as steady as an old dog.
What is MiniMax, cheap and fast, still missing?
However, judging victory or defeat only by image effects would be somewhat one-sided.
Because H3 Image has two very prominent advantages: price and speed.
Let us start with price.
According to the domestic API reference prices published by MiniMax, 1K text-to-image is about 0.143 yuan, and 2K is about 0.1531 yuan. The regular reference price for the standard version at 4K and 16MP specifications is about 0.4522 yuan, and in the first week of release a 50% discount was also offered, about 0.2261 yuan per image.
In other words, during the promotion period, a little over 20 cents can generate a high-definition image of more than 16 million pixels.
This price is indeed very competitive. Especially for creators who need to generate large amounts of e-commerce materials, advertising proposals, and game concept art, the cost advantage will be amplified as the number of generations increases.
Moreover, in this actual experience, I also clearly felt that H3 Image's generation speed was somewhat faster than GPT Image 2.6.
Of course, this is only the experience under this test entry point and settings, not a speed benchmark after strictly controlling resolution, model parameters, and server load. But for daily creation, waiting a little less means being able to try several more options, and this efficiency improvement is real.
In addition, H3 Image has opened its model weights, so developers can deploy, customize, and integrate it themselves within the permitted scope. This is also something enterprises that need autonomous control over models and workflows value greatly.
But during the experience, I also found a user experience issue that is easy to overlook: the product entry point.
Now when using GPT, whether generating images through direct conversation in ChatGPT, or creating through ChatGPT Work, Codex, and related image plugins, image capabilities are already linked with more task flows, and context can also be shared better.
When creating, we can completely discuss ideas first, organize materials, then generate images, and continue revising afterward. The whole process is not just calling an image generation model once, but continuously completing tasks within the same product ecosystem.
Although MiniMax has always adhered to a multimodal route and has laid out model areas such as text, video, image, and speech, when ordinary users actually access these capabilities, they still need to switch among different product scenarios such as Hailuo and MiniMax Code, which is very troublesome.
It also has a unified open platform that can integrate model capabilities through APIs, but for ordinary creators who do not understand development, the coherence of the front-end experience still has room for improvement, and I also hope the official team can optimize this in the future.
But that is understandable. From this actual test, H3 Image is already a competitive image generation model, especially in image realism, generation speed, and cost performance, which are indeed worthy of recognition.
But if it is to directly undertake complex commercial design, structural explanation, and precise execution tasks, GPT Image 2.6 is still more reassuring at present.
Cheapness and speed determine whether we are willing to try more, while accuracy and workflow experience determine whether it can become a daily productivity tool.