[DigitalToday reporter Yoonseo Lee (이윤서)] User complaints are spreading that the quality of Google's AI image-generation model 'Nano Banana 2' has declined compared with before.

On Aug. 5, local time, IT outlet TechRadar reported that posts have been appearing on online community Reddit claiming Nano Banana 2 generates lower-quality images than it did at launch. One user said most recently generated images look bland and that quality does not improve even with detailed prompts.

User complaints are not new. Reddit has carried posts since May saying Nano Banana 2's quality has slipped compared with the past. That is not proof of a performance drop, but it indicates some users have felt changes over time.

In early July, Google unveiled a new image model called 'Nano Banana 2 Lite.' It later integrated Nano Banana 2 into Google Earth, but temporarily halted one related feature after misuse issues emerged. There were no indications that such steps led to changes to the existing image model itself.

One possible explanation cited was a change in the text model that interprets prompts. Even if the image-generation model is unchanged, results can shift noticeably if the way instructions are read and interpreted changes. Recent changes to Gemini being in model selection and combination, rather than the image model, also aligns with that speculation.

A comparison test to gauge perceived performance declines covered three scenes. The first was a composition of looking up at the sky with a dragon flying overhead. TechRadar, which ran the test, assessed that ChatGPT's output looked natural and convincing like a real photo, while Gemini's output gave a strong impression of an 'AI-generated image.'

The second scene showed a man and a woman in their 50s talking over coffee at a cafe. Gemini's image looked awkward in character depiction and reflections, and the faces gave an impression close to the so-called 'uncanny valley.' ChatGPT's image had relatively simple detail but had fewer unnatural elements overall.

The third was an image of a British breakfast. Both models did well in food texture and plate composition, but Gemini's output showed an issue in which menu text did not look like real words. The gap was not as large as in the first two cases, but the assessment was that ChatGPT had the edge in overall completeness.

TechRadar placed more weight on the possibility that ChatGPT's image-generation ability has improved, rather than Nano Banana 2's performance having actually declined. It said OpenAI has significantly improved image-generation performance in recent months, while if Nano Banana 2 stayed at its previous level it could look relatively behind. Gemini, however, generated images faster than ChatGPT did.

As a result, the focus is shifting away from whether Google's image model itself has declined and toward Gemini's prompt-interpretation approach and the pace of improvements by competing services. Perceived image quality can also vary depending on which text model interprets instructions and how outputs are combined.

The controversy ultimately shows that not only whether Nano Banana 2 has actually declined, but also that AI image-generation quality can be judged in relative terms depending on how fast rival models advance and on users' expectations. Whether Google will further adjust its image-generation model or prompt-interpretation system, and whether it can narrow the perceived quality gap, is expected to be closely watched.