← Back to blog
AiAbout 5 min read

A domestic text-to-image model finally beats GPT on the overall chart: ArtArk takes 95.14 with a perfect consistency score

Published Nov 14, 2026
A domestic text-to-image model finally beats GPT on the overall chart: ArtArk takes 95.14 with a perfect consistency score

On November 10, SuperCLUE updated its November 2026 text-to-image benchmark. Five models were added in the supplemental run. One of them, ArtArk, from a company called 进化智能, scored 95.14 on its first appearance and took first place overall.

Second was GPT Image 2.5 Sunburst at 94.27. Third was SenseTime's SenseNova U1 Pro at 93.81.

The phrase "a domestic model on top" is worn out in Chinese tech media. Someone tops the chart every two months. So the interesting part of this news is the structure of the score table, and the ranking itself says little. ArtArk's image quality sub-score was only 87.92, the lowest tier in the top ten, almost three points under GPT Image 2.5 Flare at 90.76. Other items pulled the total back up.

What that 100 means

The benchmark splits text-to-image into six items: image quality, prompt fidelity, Chinese character generation, realistic reproduction, creation and reasoning, and complex information layout.

ArtArk took 100 on prompt fidelity. Among the 25 models entered, it was the only perfect score. Second was Flare at 93.94.

This item measures obedience. Beauty counts for nothing. The prompts carry verifiable constraints: how many people are in the frame, who stands to whose left, five red apples or one on the table, which characters appear on the sign, which direction the light comes from. The model has to land every one of them and still look natural. A perfect score means it almost never fails at that line-by-line reconciliation.

Put another way, it may not paint best. It is the least likely to improvise.

A score table from the SuperCLUE text-to-image benchmark, showing ArtArk at 95.14 overall and a full 100 for prompt fidelity.

That skill is worth little in galleries and on social feeds. It is worth a lot at the shelf. An e-commerce hero image needs the product color, the logo position and the packaging copy to be exact. A brand poster needs the slogan unchanged down to the last character. A property render needs layout, orientation and furniture count to match the drawings. These jobs used to be kept away from models, because models loved to embellish. A perfect consistency score tells the buyer you can start testing now.

The shape of the weakness

The weak side is just as clear.

Image quality at 87.92 is its lowest item and the only one below 90. That sits around the middle of the chart, which means fineness, texture and light grading trail the top models by a visible margin.

Chinese character generation at 97.02, slightly under Sunburst at 97.25 but close. Realistic reproduction at 96.02, also slightly under Sunburst at 96.49. Creation and reasoning at 96.47, under GPT-Image-2 at 96.81. Complex information layout at 93.38, more than two points behind SenseNova U1 Pro at 95.80.

So it leads across the board on execution and trails across the board on aesthetics. This is a specialist's profile, and an honest one.

The chart itself does not hide it. It plots the 25 models on a four-quadrant map, scenario application on the x axis and general capability on the y. ArtArk lands with Sunburst, U1 Pro and Qwen-Image-3.0-Pro in the "outstanding leaders" cell, but toward the lower right, farther from the "technology leader" corner.

The top five differ by 1.91 points, and that is the harder problem

Sunburst 94.27, ArtArk 95.14, a gap of 0.87. Across the whole top five, the highest and lowest are 1.91 apart.

Three years ago, charts like this routinely showed a dozen points between the head and the middle, and first place carried no suspense. The gap has now dropped into the noise range of a single evaluation. The chart itself concedes that domestic models within one point count as tied. By that rule, ArtArk, Sunburst and U1 Pro belong to one tier.

Smaller gaps create a practical problem. The total score tells a buyer nothing. You have to read the sub-scores.

This round the sub-score leaders are scattered. The top marks for Chinese character generation and realistic reproduction both sit with Sunburst. Complex information layout goes to SenseNova U1 Pro at 95.80. Creation and reasoning goes to the previous generation GPT-Image-2 at 96.81, higher than both new versions this year. Prompt fidelity belongs to ArtArk.

Four different "strongest" titles on one chart means this generation is no longer climbing the same ladder. The routes are splitting fast: some are filling in semantic execution, some image texture, some layout organization. Model selection is turning from "pick the best one" into "pick the one that matches".

How fast the versions move

The Nano Banana line deserves its own look. Version 2.1 is 2.74 above version 2, with prompt fidelity up 9.09 and complex information layout up 6.87. At the same time its image quality, realistic reproduction, creation and reasoning barely moved.

That curve is unusual. A normal generational change lifts every item a little. One item up nine points while two stay flat means a targeted strike, aimed at the previous generation's weakest semantic alignment. Google evidently read the same evaluation report.

The other side of it: ArtArk's structure, maximum consistency with bottom-tier quality, carries risk. Consistency can be closed quickly with engineering and training data. Texture cannot. Texture rests on aesthetic judgment and data distribution, and it grinds slowly. Someone could catch up in a single quarter.

One-sentence conclusion

The real information sits behind the ranks. The center of gravity in text-to-image evaluation has moved. Making a picture look good is something the top twenty-odd models all clear. Whether a model can strictly execute one complex sentence is becoming the dividing line.

For people who make images, the working method has to change. It used to be wishing: pull ten, pick one. Now you can write a brief, state the constraints the way you place an order, then inspect the delivery. The more obedient the model, the less effort goes into prompts and the more into judgment. When every model can deliver "good looking", only you know which frame is right.

Related articles