I’m sorry, but nano banana is just imagen 4 with the ability to intake visual reference, it doesn’t mean it’s good at it, nor is it good at keeping the art styles or even making a not conventionally neat looking image, this model only able to accessed through gemini 2.0 flash was leagues ahead of other image editing models, it’s accuracy at preserving the image you sent it and the accuracy of it’s understanding of prompts was, and arguably is still unmatched compared to the new “nano-banana” model, nano-banana is just not good enough to be considered a replacement for this model, and it should have had a proper release so it could be accessed beyond gemini 2.0 flash.
Bring back the unnamed image generation model that was previously available through Gemini 2.0 Flash
I should probably leave at least one example.
when it was capable of intaking more than one image, gemini 2.0 flash image preview was more capable than nano-banana at retaining very distinctive details without needing to tell it specifically, it was capable of making images that didn’t have that standard ai generated look to them which made it more versatile in photoshopping than the new model. It’s only flaw was the fact it was a much smaller model than nano-banana but it was still very impressive for what it was. But after receiving multiple “updates” the model became less capable, if I prompted this in June I would have gotten something completely unrelated to what I asked for.
Nano-banana also struggles with art styles, I’ll use the same set of images as examples.
It’s clear that nano-banana did not recognize both images as the same subject and only copy and pasted one image, while 2.0 flash created an original image that resembles the subject, easily avoiding producing the same image, the other issue is with replicating styles, you can see how the 2.0 flash understand fully how art works, the image looks hand drawn, from first glance you wouldn’t be able to tell the image is ai generated if it wasn’t for the watermark and the image isn’t visually identical to the reference on top of that, nano-banana does not even take one of the references into account and simply reproduces the same image from a slightly different angle, even though it’s generating off of one image the subject even looks slightly different, and prompting does not solve this issue at, it doesn’t matter how you word it, it will never produce anything on par with the creativity of 2.0 flash.
I really think retiring gemini 2.0 flash’s image generation model should be reconsidered, the model was constantly nerfed with each monthly update to it, When it initially launched it was genuinely impressive, it became more buggy and less functional which is speculated to prioritize gemini 2.5, If this model’s availability wasn’t directly to the Gemini API, it would have a different impression on people.