The Character Consistency is Bananas

Gemini's latest image model is crushing charts

Google Deepmind just released its newest image editing model, affectionately known as Nano Banana (officially named Gemini Flash 2.5). Within hours, it's shot to the top-rated image editing model in the world.

And for good reason. For a long time, image generation models have struggled with character consistency and editing. Send a reference photo of a person, and the models would return something that “looks close but not quite”. Or try to edit a part, and it would hallucinate. e.g. remove a chair and suddenly a lemon appears.

Nano Banana seems to have cracked much of that. I ran a series of tests to see how consistent it really is (p.s. scroll to the bottom for my Founder & Builder takeaways):

Reference: Here’s the baseline image I used.

New environment: Now she’s at a grocery store. It’s incredible how it preserved the detail of the jacket.

Selfie angle: With a different perspective. Note how it captured the reflectiveness of the sequins under different lighting.

Outfit swap: Easy change of outfits.

Style transfer: Turned the photo into an illustration, respecting the reference style.

Stress test:Asking it to put me in a different outfit and in a different environment seemed to override its consistency. The likeness didn’t hold!

Multiple people: It also does somewhat “ok” with multiple people. Here’s me with my friend and collaborator Burton Rast. Close enough, but a bit uncanny.

Takeaways For Founders and Builders:

• Image generation + editing just leveled up. With Gemini’s Nano Banana, you can now use these tools through the API in production, not just tinkering. Think programmatic ads, personalized album covers, sales decks tailored to each prospect. Visual customization is now a reliable building block.
• Branches unlock consistency at scale. This under-discussed feature lets you spin out coherent variations: a sequence of stills stitched into stop-motion, a product catalog where every image feels part of the same shoot, or even early video workflows. It’s the bridge from one-off novelty to repeatable creative pipelines.