AI News

Hands-on Test of MiniMax H3-Image, GPT Image 2.6 and Seedream 5.5: Who Is Faster and Who Is More Accurate

TMTPost · AGIUpdated Nov 14 10:06

AI summary

The author tested three new image models—MiniMax H3-Image, OpenAI GPT Image 2.6 and ByteDance Seedream 5.5—via Hailuo AI, ChatGPT and Doubao. MiniMax generated images in 10-15 seconds in Hailuo AI's high-speed mode, leading on speed, but under complex prompts it showed clipping and unfused faces; OpenAI ranked first on both the Artificial Analysis text-to-image and image-editing leaderboards, with the most accurate understanding and modification of complex instructions; Seedream 5.5 missed changing one face during local editing, and a second revision took about 7 minutes. The article also analyzes the three companies' strategies: OpenAI relies on the ChatGPT entry point, ByteDance relies on content scenarios such as Jimeng, and MiniMax seeks cost-sensitive customers with speed, low prices and open weights.

Why read this

AI score 58

The author used the same set of complex prompts to compare three new image models side by side, revealing the actual gaps in speed, instruction comprehension and editing stability.

Topics

Tags

Body · AI translation

By AIX Finance, Author | Lei Jing, Editor | Wei Jia

On the evening of November 10, MiniMax released and opened the model weights of H3-Image, playing the "fast and cheap" card. The company claims it can generate a 16-megapixel high-definition image in 10 seconds, at only 50%-70% of the price of mainstream tools in the same tier.

Before it, OpenAI and ByteDance had just handed in their homework.

On November 5, OpenAI launched Images 2.6GPT Image 2.6, continuing to push on speed and editing capabilities, claiming a 50% improvement in image generation speed over the previous generation, adding GIF animation generation and reusable reference images, and launching two models on the API side, GPT Image 2.6 Flare and Sunburst, the latter focusing on consistency in multi-turn editing.

ByteDance, meanwhile, released Seedream 5.5 on November 9, focusing on the stability of continuous editing and supporting precise local modifications.

All three have rolled out their own new models. How big is the actual performance gap? Beyond competing on generation quality, what else are they fighting over?

01. OpenAI leads, MiniMax wins on speed

Let's look at the leaderboards first—the gap between the three is quite large.

In Artificial Analysis's latest text-to-image leaderboard, GPT Image 2.6 Sunburst (max) ranks first, Seedream 5.5 ranks tenth, and H3-Image ranks twenty-third.

Image source / Artificial Analysis official website

On the image editing leaderboard, OpenAI still holds first place, ByteDance rises to sixth, and MiniMax climbs to twelfth. Compared with generating images from scratch, the latter two rank somewhat higher in editing capability.

Image source / Artificial Analysis official website

OpenAI, having taken two first places in a row, can it really leave the other two behind in actual use?

We experienced the three new models through ChatGPT, Doubao, and Hailuo AI respectively. Among them, ChatGPT used GPT Image 2.6 automatic mode, Doubao called Seedream 5.5, and Hailuo AI tested both the standard and Highspeed modes of H3-Image.

It should be noted that OpenAI's GPT Image 2.6 offers different model tiers via API, and developers can choose as needed. But most ordinary users generate images directly in ChatGPT, so we did not call the API separately and tested the actual experience on each product side.

The first task focuses on comprehension.

There's no need to elaborate on how popular short dramas are. First there were the fruit ethics dramas of pineapple betraying strawberry and banana cheating on grape, then there were the melodramatic short dramas of wood ear mushroom and manicure competing to be the heir of a wealthy family. Short drama protagonists are no longer limited to humans—even food has started dating and fighting over family fortunes.

Recently, the wildly popular extra-long-name tanghulu naturally hasn't escaped either. With a long list of ingredients comparable to chanting a spell, it successfully infiltrated high-society circles. In its family, the more ingredients and the longer the name, the nobler the bloodline; ordinary hawthorn tanghulu doesn't even qualify to inherit the family fortune.

English text is being prepared.

We also specifically emphasized that all characters have human bodies and tanghulu heads, the ingredients must blend naturally with the facial features, and no ordinary human faces may appear.

More than a dozen ingredients, multiple characters, complex settings—one task tests comprehension, execution, and visual expressiveness at the same time.

Prompt:

Theme: The extra-long-name tanghulu family, the battle over the right to inherit a ten-billion fortune

Generate a cinematic 3D animation still from an absurd Chinese AI short drama, aspect ratio 9:16, exquisite semi-realistic 3D animation style, with peculiar, exaggerated character designs and a strong comedic effect.

Core setting: a top-tier wealthy family made up of extra-long-name internet-famous tanghulu. All characters have normal human bodies, wearing luxurious suits and gowns, but their entire heads are giant three-dimensional tanghulu stacked from more than a dozen different dessert ingredients.

Note that these tanghulu are not traditional skewers of a few hawthorns, but are jointly composed of ingredients such as hawthorn, strawberry, Shine Muscat grape, blueberry, mango, milk skin, taro paste, blood glutinous rice, mochi, cheese, mango pomelo sago, crushed pistachio, chocolate crispy shell, cold brew yogurt, and butter crispy flakes.

Different ingredients are layered one on top of another: bright red hawthorn and strawberry, emerald green grapes, golden yellow mango, milky white thick milk skin, light purple taro paste, deep purple blood glutinous rice, soft white mochi, dark brown chocolate, and golden yellow butter crispy flakes interwoven with each other; some fruits are cut open, revealing filled cheese and yogurt.

The entire head is like a giant, colorful dessert pagoda, clearly extending upward and much taller than a normal human head. The outer layer is wrapped in crystal-clear transparent sugar coating, with a slender bamboo skewer exposed at the top, and the syrup glistening under warm lighting. Ingredient details are realistically discernible, with obvious thickness and three-dimensional structure between layers.

The tanghulu heads have eyes, eyebrows, and mouths naturally integrated into the food shapes, able to express rich emotions such as anger, arrogance, and shock, without appearing as ordinary human faces.

Scene: an extremely luxurious private banquet hall of a wealthy family. A huge golden crystal chandelier, a high ceiling, white marble floors, and a full wall of floor-to-ceiling glass windows. Outside the windows is a bustling city night street with flickering neon lights.

In the center of the banquet hall is a long marble dining table, set with tall red wine glasses, golden candlesticks, and exquisite tableware.

Characters and plot: On the left side of the image is an imposing elderly male patriarch of the family, wearing a premium dark brown suit, a gold ring, and a family crest on his chest. His head is an extremely enormous, super-luxurious tanghulu made of more than twenty layers of different fruits and desserts, with extremely complex ingredients stacked layer upon layer upward, forming an exaggerated crown-like shape. He has an angry expression and extends his right hand, pointing toward the other side of the dining table.

Beside him stands a family matriarch, wearing a creamy white silk gown, her head formed by layers of green grapes, milk skin, pistachio cheese, white mochi, green matcha crispy shell, and a golden sugar shell, creating an elegant and complex green-and-cream tanghulu tower.

On the right side of the image is a young heir of the wealthy family, wearing a finely tailored white-and-gold suit, standing tall and upright, confronting his elder with his head held high. His head is a huge colorful tanghulu composed of strawberries, hawthorn cakes, mango pomelo sago cheese, taro paste, blood glutinous rice, mochi, milk skin, pistachio cold brew yogurt, and chocolate butter crispy flakes, with complex layers and an appearance like an exquisite dessert tower of more than ten layers.

Next to the heir is a young woman, wearing a light green high-end gown, with an ordinary hawthorn tanghulu for a head, and a shocked expression.

Composition and style: Use a symmetrical panoramic composition, with the dining table in the center of the image and the four characters arranged on both sides, with all characters fully in frame from head to toe. The tanghulu heads are tall and eye-catching, occupying a significant visual proportion of the image, but must not obscure the other characters. The characters' body proportions are natural, their clothing is gorgeous and exquisite, and the tanghulu heads are exaggerated yet realistic, creating a strong contrast.

Cinematic 3D animation rendering, exquisite food materials, crystal-clear sugar shells, the crispy texture of milk skin, the dense texture of cheese, the soft texture of mochi, the vivid colors of fruits, and rich highlights and reflections.

First, speed. MiniMax is indeed fast. However, we almost missed it.

Initially, when we called H3-Image through MiniMax Code, we found that generating one image took about 1 minute, and a single edit even took 5 minutes. What happened to the promised 10-second image generation?

Later, we switched to Hailuo AI, and the speed finally picked up. H3-Image in standard mode takes about 25-30 seconds per image, while high-speed mode takes only 10-15 seconds. The reason is that MiniMax Code, as an Agent product, involves not only calling the image model but also task scheduling and other steps, and the entire process incurs extra time.

By comparison, OpenAI and ByteDance models both take about 1 minute per image generation. On speed, MiniMax leads.

Now let's look at the results. ChatGPT clearly understands this kind of abstract art. After more than a minute, a still from a wealthy-family short drama was produced. The characters are gorgeous, the expressions are exaggerated, and the tanghulu and facial features are blended quite naturally. Luxurious and dramatic, as if the next second someone will slam the table and throw out a check: "Here are five million. Leave my son."

Hailuo AI generates 4 images at a time. The image quality difference between standard mode and high-speed mode is not significant, but the mistakes they make are the same. The first image in both showed obvious clipping, with characters standing directly on the table, turning a wealthy-family meeting-the-parents scene into the wealthy family collectively "getting on the table."

Image generated by MiniMax H3 standard version

The remaining three images, although without obvious clipping, still did not get the heads right. Everyone had an ordinary human face, with a large skewer of tanghulu standing on top of their heads, as if wearing a deluxe tanghulu hat. Some characters' expressions were also rather flat, lacking a bit of dramatic flair.

We selected the image with the biggest problems as a reference and asked Hailuo AI to put the characters standing on the table back down and then blend the human faces with the tanghulu. The four regenerated images became a collective "getting on the table," all showing clipping problems, with even more characters on the table, as if performing some mysterious ritual. As for the faces and tanghulu, they still grew separately.

Image after modification by MiniMax H3 standard version

Doubao also stumbled on the heads. There were no major problems with the scene or clothing, but the characters likewise had ordinary human faces, and the tanghulu and facial features also grew separately.

Still generated by Doubao

Fortunately, it supports local editing. We selected the characters and asked it to change the heads, but out of four faces it only changed three, and one character continued to have a human face, mixed into the tanghulu family.

We had to remind it again. This time we waited about 7 minutes, and finally dealt with the one that slipped through. But after the modification, although the food and human faces were barely blended together, the characters' original expressions and identity features became blurred, and the young master and his father looked as if they had swapped identities. This is still a long way from what Seedream 5.5 claimed about "stable multi-round editing"!

Next is the second test, which examines aesthetics.

Previously, a Chinese heavenly palace panorama went viral online: the heavenly court above a sea of clouds, with palaces, immortal mountains, and white jade heavenly stairs all present, ethereal and otherworldly. This time, without complex characters, we only asked the three AIs to create a magnificent mythological heavenly court, to see whose aesthetics are more on point.

Original Chinese mythological heavenly court panorama, a grand Eastern wonderland on an almost cosmic scale. The camera is positioned above the ninth heaven, using an ultra-wide-angle high-altitude overhead view to reveal a heavenly city suspended in a boundless sea of clouds. Hundreds of Chinese classical palaces, jeweled towers and jade pavilions, and vermilion towers and pavilions spread out in layers upon giant floating immortal mountains; the palace gates and watchtowers are distributed in a staggered manner along the mountain contours, extending toward the distant horizon and presenting a magnificent yet orderly architectural complex.

At the center of the image is the Lingxiao Treasure Hall, built on a giant white jade platform, with a double-eaved hip roof, golden glazed tiles, giant vermilion columns, gilded bracket sets, and exquisite dragon-and-phoenix patterns carved on the roof ridges. Surrounding the main hall are dozens of immortal palaces of varying sizes; white jade long bridges span the sea of clouds, connecting the floating peaks. Thousands of heavenly steps spiral upward along the mountain ridges, finally converging at the Lingxiao Treasure Hall at the highest point. The roads, stairs, and bridges between the palaces have reasonable connecting relationships, and every building has a clear supporting structure and realistic proportions of Chinese classical architecture.

On the left side of the image, a giant floating immortal mountain rises from the clouds; vigorous ancient pines grow on the cliffs, dotted with pavilions and towers roofed in blue-green glazed tiles. On the right is a grand heavenly palace complex, where layered palace walls, white jade courtyards, and golden flying eaves interweave and extend in the sunlight. Several giant waterfalls pour down from the suspended cliffs and fall into the bottomless sea of clouds below. In the distance, three-tiered immortal mountains loom faintly in the morning mist, and countless tiny palace silhouettes appear on the horizon, creating extremely spectacular spatial depth.

The sky presents a natural gradient of pale cyan, azure, and gold. Morning sunlight shines from the east, piercing through the layered sea of clouds and forming clear volumetric light among the immortal palace roofs, white jade long bridges, and mountain ranges. The sea of clouds has a realistic sense of flow, thickness, and natural occlusion relationships. Nearby buildings have fine textures, while distant palaces gradually become smaller and paler with distance, forming accurate atmospheric perspective.

It blends the mood and negative space of Chinese classical landscape painting, the majestic composition of Song dynasty landscape painting, the refined details of fine-line heavy-color painting, and the realistic visual effects of top-tier Eastern mythological films. The architecture follows traditional Chinese palace forms; realistic flying eaves, bracket sets, mortise-and-tenon joints, roof-ridge divine beasts, stone-carved balustrades, dragon patterns, and cloud-thunder patterns are clearly visible. The overall palette is dominated by white jade, vermilion, gilded gold, blue-green, and pale cyan; the colors are bright, translucent, rich yet restrained.

The image is sacred, solemn, tranquil, and magnificent, with a strong sense of architectural gigantism and epic scale. The emphasis is on highlighting the grand volume of the heavenly court complex, its spatial order, and the hierarchical relationship between heaven and earth. No people, immortals, heavenly soldiers, or humanoid creatures appear. Cinematic digital concept art, ultra-wide-angle high-altitude overhead view, magnificent deep composition, 16:9 horizontal composition, ultra-high detail, 8K visual quality.

Negative prompts: people, humanoid statues, Western Gothic churches, Greek temples, European castles, European angels, modern cities, glass curtain walls, science-fiction architecture, cyberpunk neon, spaceships, cartoon style, plastic texture, incorrect architectural structure, imbalanced roof proportions, repeatedly duplicated pavilions, unsupported bridges, chaotic floating relationships, broken stairs, perspective errors, overly crowded buildings, thick fog obscuring the subject, excessive golden light, excessive purple, random text, Logo, watermark, existing film and television characters, game characters, and IP symbols.

ChatGPT continues on the refined route. The scale of the heavenly palace is quite astonishing; palaces densely fill the image, architectural details are maxed out, and the mist-shrouded atmosphere is also in place.

It's just that it seems to have its own obsession with the word "magnificent." The entire heavenly court is resplendent in gold and jade, with so many palaces that they almost cannot fit; it seems desperate to cram every real-estate project of the ninth heaven into one image, and the more you look, the more it feels a bit nouveau riche. I wonder if it's because it has been in first place for so long that even building a heavenly palace, it doesn't know how to keep a low profile.

It's a good opportunity to test its editing ability. We directly pointed out the problems in the image and asked it to reduce the number of buildings, tone down the opulence, and add some Chinese charm. After about 1 minute, the revised version came out: fewer palaces, a cleaner image, and the style was right. Its willingness to take advice: full marks.

Hailuo AI's heavenly palace layout is relatively more spacious, with clouds and mist weaving among the pavilions, like a well-made xianxia game scene image. Moreover, whether in standard mode or high-speed mode, the four generated images are fairly consistent in style, with the differences mainly in building arrangement.

Doubao's performance is very close to it: palaces advance layer by layer, the composition is natural, and it likewise has a strong xianxia flavor.

After two rounds of testing, OpenAI's model still leads in overall performance, understanding complex instructions most accurately and making revisions quickly and accurately, basically getting it right in one go, though its aesthetics occasionally overdo it. MiniMax's model showed a speed advantage, producing images on Hailuo AI in as fast as a dozen or so seconds, and the image quality is not bad either, but when facing complex tasks, it has not yet achieved both speed and quality. ByteDance's model has good image quality, but in this hands-on test, revisions were both slow and not stable enough, which disappointed us somewhat.

02. Models are getting stronger, and vendors with scenarios are more likely to win

Vendors continue to invest in image model research and development because they see the commercialization space in the professional market.

Compared with entertainment needs such as generating avatars and emoji packs, industries such as advertising, e-commerce, and game art need to continuously produce large numbers of images and are also more willing to pay for efficiency. As model capabilities improve, AI has the opportunity to take on more design and asset production work, helping companies save labor costs.

There is more than one way to monetize image models. In addition to charging individual users membership fees, vendors can also earn revenue through API calls and enterprise services, and can integrate image capabilities into existing products to increase user usage frequency.

But professional users also have higher demands. Images not only have to look good, but also have to ensure consistency of products, people, and brand elements, and be accurately modified as required. From the tests above, the current models still have quite a few problems to solve.

Facing the same market, the three companies have different approaches.

OpenAI competes on model capability and entry points. ChatGPT has a huge user base; users do not need to open dedicated image-generation software, and can generate and modify images right in the chat box. OpenAI disclosed that users worldwide generate 4 billion images per week on ChatGPT, Codex, and the API. Every time it upgrades its image model, OpenAI can directly reach this group of users without promoting a new product from scratch. Conversely, image capabilities can also increase ChatGPT usage frequency and attract users to pay.

ByteDance, meanwhile, has more mature content creation scenarios.

Seedream can be integrated into products such as Jimeng, Doubao, and Doubao Work, and can also provide services to enterprises and developers through Volcano Ark. Ordinary users can use it to generate fun images, professional creators can produce assets, and enterprises can also generate content in batches through the API.

Image source / Screenshot from the Seedream official website

Jimeng in particular already gathers a large number of users with image and video creation needs, and every Seedream upgrade can directly serve these needs. ByteDance can also collect usage feedback through existing products to continue improving the model.

MiniMax is in a relatively more difficult position. Although it has creation products such as Hailuo AI, its user scale and scenario maturity are far below those of ChatGPT and Jimeng. Even if its model capability is good, getting users to switch to a different tool requires extra customer acquisition costs.

Therefore, MiniMax chose to seek breakthroughs in speed, price, and an open strategy.

In this test, the image quality of H3-Image's high-speed mode showed no obvious gap with the standard mode. For users who need to generate images in batches, faster speed means higher efficiency. Coupled with relatively low generation prices, MiniMax has a chance to attract cost-sensitive developers and enterprise customers.

But speed and low prices are unlikely to form a long-term barrier. OpenAI and ByteDance can likewise increase speed and cut prices, and vendors with larger user bases are more likely to dilute costs through economies of scale. What professional users care about more is whether an image can be modified correctly in one pass as required. If an image has to be revised several rounds, the few dozen seconds saved by faster single-image generation are not of much significance.

Opening up model weights has greater long-term value for MiniMax. Enterprises can carry out local deployment and customization according to their own needs, reducing dependence on third-party API services. Compared with closed-source models that can only be called through APIs, this is more attractive to enterprises and developers with customization needs. Once an enterprise completes deployment and system integration around a model, switching models entails extra costs, giving MiniMax a better chance to retain customers.

MiniMax has proven that it can generate images faster and more cheaply, but in the next test, we hope it can not only generate images in a dozen or so seconds, but also get the heads of the Tanghulu family right in one pass.

Source

TMTPost · AGI