Descriptions:
Dan Kieft runs Nano Banana 2 — described in the video as officially named Gemini 3.1 Flash — through an extensive series of capability tests inside the OpenArt platform, which allows direct side-by-side comparison against Nano Banana 1, Nano Banana Pro, and Cadream 5.0. Tests span historical scene reconstruction using GPS coordinates of the Roman Colosseum across four time periods (80 AD, 1450, 1870, and 2025), photorealistic portrait generation, image-to-image transformation, multilingual text rendering, and multi-subject consistency with up to 14 simultaneous objects and characters in a single image.
On historical world knowledge and period-accurate architecture, Nano Banana 2 outperforms all compared models, though it still falls short of perfect accuracy in specific details. Typography is highlighted as a genuine differentiator — the model correctly renders and translates text from a German newspaper and generates an accurate Japanese billboard. The multi-subject consistency test combines 14 reference images (characters, objects, backgrounds) into a unified animated movie poster using OpenArt’s auto-polish feature and per-image tagging, producing a result Kieft describes as commercially usable.
Practical techniques covered include using OpenArt’s character tagging system to disambiguate which input image corresponds to each described element, leveraging auto-polish for complex multi-element prompts, and pairing Nano Banana 2 with Higgsfield Soul for layered character-consistent image workflows. The video is aimed at image creators weighing where this model fits relative to the current generation of image generation tools.
📺 Source: Dan Kieft · Published March 09, 2026
🏷️ Format: Review
