Early look at images generated by Nano Banana 2 via Media AI (www.testingcatalog.com)

Google appears to be gearing up to release Nano Banana 2 (seen in a recent Media AI preview), a substantive upgrade to the image-generation model used across Gemini and related products. Leaked/preview outputs show big gains in tasks that Nano Banana 1 struggled with: precise coloring, fine-grained control over view/angle, and correcting or rendering text. Internals indicate the model currently runs on Gemini 2.5 Flash (not yet Gemini 3.0 Pro), and introduces a new multi-step generation pipeline that plans an image, generates it, performs built-in image analysis to spot errors, and iterates corrections before returning the final output — an iterative loop that improves fidelity and consistency across sessions. Technically, code references and tests point to expanded aspect-ratio support (1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9, 21:9) and 1K/2K/4K resolution modes. Internal rebranding to “Nano Banana Pro” and claims of up to threefold gains in instruction-following accuracy suggest this build targets professional workflows (reconstruction stress tests like “shredded” image recovery were highlighted). Uncertainty remains about the ultimate base (Imagen 4 vs. Gemini 3 Pro/Flash) and exact rollout scope, but announcement cards in Gemini and multi-product plans (e.g., Whisk labs) signal a public release is imminent — a meaningful step for creatives and product teams seeking higher-accuracy image generation and tighter editorial controls.