[DigitalToday reporter Jinju Hong] OpenAI’s new image generation model, ChatGPT Image 2.5, has improved an over-sharpening issue flagged in the previous version, but it was hard to pick a winner in a comparison with Google-side Nano Banana 2. Across 6 evaluation categories, the two models split the results 3-3, producing a near draw.

On Sept. 12 (local time), blockchain outlet Decrypt reported that in the comparison test ChatGPT Image 2.5 showed strengths in spatial composition and mood in complex scenes, as well as in illustrations. Nano Banana 2, by contrast, more reliably handled verifiable details such as text and dates within images.

OpenAI, when releasing ChatGPT Image 2.5 on Sept. 8, highlighted sharper detail, richer texture, more natural lighting and precise editing functions as key strengths. It also improved the ability to edit images while keeping elements that users specify must not be changed. Image generation latency is down by up to 50 percent compared with Image 2.0, it said.

Two models were added to the API: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Flare focuses on basic high-speed image generation, while Sunburst is a higher-tier model aimed at precise image editing.

This test follows a comparative evaluation conducted in May. At the time, OpenAI’s model was criticised for producing overly sharp images when handling prompts containing multiple constraints.

In this test, that issue did not appear to a significant extent. In steampunk scenes and portrait photos, ChatGPT Image 2.5 produced complex scenes relatively stably without a strong yellow cast or a sense of heavy post-processing.

Still, there were scenes where Nano Banana 2 led on detailed accuracy. In particular, in an image of a city intersection containing a large amount of text, Nano Banana 2 rendered most phrases relatively clearly.

By contrast, ChatGPT Image 2.5 did not properly render the apostrophe in “KELLERMAN'S” or made it hard to read, and it misspelled some words in street-art phrases. Overall scene realism was high, but Nano Banana 2 had the edge in text accuracy within images.

ChatGPT Image 2.5 showed strength in spatial composition and mood. In the clock-tower category, it rendered steam, a river and tonal differences driven by perspective naturally, producing a more three-dimensional image.

The evaluation said ChatGPT Image 2.5 followed the given instructions relatively faithfully while maintaining realism. Nano Banana 2 made Roman numerals easier to read, but the time shown by the clock hands did not exactly match the prompt.

ChatGPT Image 2.5 also received favourable marks in scenes with a strong illustrative character. In an animation-style spiritual transformation scene, it was rated highly for its depiction of the sky and overall visual completeness. Still, it had limitations in rendering detailed instructions literally. Rather than directly depicting a person melting into energy, it interpreted the instruction by adding an electric effect around the hair. Nano Banana 2 reflected the instruction more literally, but ChatGPT Image 2.5 led in overall scene completeness.

In realistic portrait photos, the two models’ strengths and weaknesses diverged. In a category depicting an architect on a rooftop, ChatGPT Image 2.5 rendered lighting and skin texture well, but there was an issue of the skin appearing overly smooth.

Nano Banana 2 did not place the blueprint in the correct position, but it was relatively stable in overall composition and in rendering document labels. An evaluation also said Nano Banana 2 was competitive in a single generated image, but ChatGPT Image 2.5’s strengths stood out when repeated revision work was included.

The biggest difference appeared in research-based images requiring fact-checking. Both models support making images after researching materials, but in a Bitcoin timeline infographic ChatGPT Image 2.5 mislabelled a key year. ChatGPT Image 2.5 marked the approval timing for U.S. spot Bitcoin exchange-traded funds (ETFs) as 2023. The first spot Bitcoin ETF in the United States was approved on Jan. 10, 2024.

Nano Banana 2 also did not produce a perfectly organised structure for the infographic, but it avoided a definitive error by grouping the timing of Bitcoin ETF approval and the fourth halving as 2023 to 2024.

The evaluation pointed to such confident but incorrect dates as a more serious flaw than a simple design issue. It showed that when generative AI produces fact-based content, accuracy of information can be a core evaluation factor, not only image quality or design.

In categories that visualised abstract or meaningless coined words, ChatGPT Image 2.5 had the advantage. ChatGPT Image 2.5 directly rendered words with no fixed meaning as text on signs and on a board-game box. The evaluation cited as a strength its conversion of undefined concepts into visual elements that can actually be read. Nano Banana 2 created a warm, cultural atmosphere, but it did not explicitly surface the coined word from the prompt within the scene.

OpenAI also strengthened related functions beyond the image model itself. It added a sketch function that lets users draw a rough composition directly within ChatGPT, a prompt-sharing function and an inline comment function that lets users leave comments on specific image areas. It also provides format-based templates that can be used for posters and product creation. The API’s quality tiers were also expanded, with new “high” and “max” levels added.

Overall, the match-up was effectively a near draw. ChatGPT Image 2.5 fixed a representative flaw of the prior generation and produced some of the most impressive results in certain scenes. Nano Banana 2, by contrast, was more stable on verifiable detail accuracy such as spelling, dates and notation. A key takeaway from the comparison was that the gap came down less to overall image quality than to small, checkable mistakes such as text errors and incorrect years.