Google's Nano Banana 2.1 Holds Its Own Against ChatGPT in Image Generator Tests

The upgrade promises sharper details, better instruction-following, more accurate text and greater consistency across images. TechRadar said Gemini accounts should have been upgraded to Nano Banana 2.1 by now. The test gave Gemini and ChatGPT identical prompts designed to assess different abilities and compared the images side by side, with ChatGPT on the left and Gemini on the right. OpenAI's Images 2.5 has made progress in realistic lighting, complicated compositions and text rendering, making image generation one of ChatGPT's most entertaining and useful features, the report said.

The first prompt asked for a photorealistic image of a tiny Victorian greenhouse inside a glass snow globe during a thunderstorm, with realistic reflections, refractions and lighting. Both generators produced images that would suit a fantasy novel cover. ChatGPT made an elaborate, golden-lit Victorian greenhouse inside a large glass sphere on a polished wooden base beside a rain-covered window. Lightning lit the storm outside, while the globe held a detailed miniature world with plants, pathways and delicate architectural flourishes.

Nano Banana 2.1 took a more restrained approach. Its greenhouse looked more like an actual miniature structure, with dark metal framing, lush greenery and warm interior lighting that contrasted with the blue-gray storm outside. The globe's reflections were particularly convincing, and the lighting felt a little more natural, TechRadar said. The report noted an odd choice: a wet tabletop even though the scene appeared to be indoors.

The second test examined text rendering, a long-standing weakness for AI image generators. The prompt asked for a retro-futuristic travel poster for Saturn's moon Titan with the exact text 'TITAN IS CALLING,' 'Seven nights beneath the rings,' and 'Starting at $2,499.' ChatGPT created a luxurious travel advertisement with Saturn dominating the sky above an orange landscape, futuristic resorts and a couple enjoying drinks, suggesting Titan as an exclusive vacation destination. TechRadar disliked the gap ChatGPT left between the dollar sign and the number 2.

Google's version was more cartoonish, a colorful vintage science-fiction poster with rockets, astronauts, icy mountains and a futuristic base. TechRadar said it was more realistic about the moon's inhospitable conditions but still had the optimistic quality of a 1950s space-travel advertisement, with bold lettering and a slightly weathered finish. Crucially, it reproduced the requested text correctly, including the price.

A third test focused on character consistency, another tricky problem for AI image generators when the same character needs to appear in several scenes. TechRadar noted that a fox might suddenly change appearance, lose an accessory or acquire an extra limb between panels. The prompt asked for a four-panel comic showing an orange fox wearing a green scarf discovering a magical key, unlocking a door and entering a library, with the fox kept identical in every panel.