Gemini 3.1 Flash Image

Gemini 3.1 Flash Image — codename "Nano Banana 2" — is Google DeepMind's Gemini-native image generator. Google describes it as natively multimodal: text, image, audio and video share one token space and one transformer, rather than an image pipeline bolted onto the LLM, with diffusion-style iterative denoising happening inside that shared architecture. Secondary technical analysis calls it a "Multimodal Diffusion Transformer"; that has not been confirmed against a Google-published architecture paper, so the `diffusion-transformer` classification here is the closest defensible match, not a verified specific. It is distinct from Google's separate Imagen line, a more conventional dedicated diffusion model.

The "Flash" tier is positioned for speed alongside "Pro-level" quality claims — the speed half being the more verifiable of the two. Per Google DeepMind's own model page it renders legible in-image text with font and style control, does in-image translation and localization, grounds subject accuracy in Gemini's knowledge base and web search rather than prompt parsing alone, upscales output to 2K/4K, and holds up to 5 characters and 14 objects consistent across a series of generations.

There is no independently published GenEval or T2I-CompBench score for it, so the capability ratings in this entry are vendor-stated. A press claim of a #1 Text-to-Image ranking was checked directly against Artificial Analysis's own leaderboard (2026-08-19): Nano Banana 2 sits at #3 of about 147 models tracked (Elo 1320) — strong, but behind GPT Image 2 (high) at Elo 1368. Modelglass records this entry's capability confidence as low.