Gemini-2.0-Flash-Preview-Image-Generation Has Odd Behaviour

So Google just recently deployed for use in the Vertex AI API as well as the Vertex AI Studio an enhanced model aka the successor to the experimental model they’ve been offering for a while now (the one in the title of this post).

i keep coming across these situations, where I submit to the model an image of a room (bedroom for instance) followed by an instruction to empty the room (make an edit/modification to the image) and the model rambles on for quite some time, almost like it’s ouputting some chain-of-though output, giving back more and more images (each newly generated image is more distorted and nonsensical than the last).

take this example. i gave it an image of a decorated bedroom and it kept returning images, one after the other, like the ones I’ve attached (image1 followed by image2 followed by image3)



the prompt I used was: “generate an image of this bedroom with all bedroom decoration removed. do not alter the room structure or architecture. please only generate one image.”

Hi @GerardT,

Thanks for reporting. Looks like some model quality related issue. I will escalate with our engineering team.

Its really bad like worst the face doesnt look like a face. Every generation after the update, not even a single image is usable. Can give you specific examples if that would be helpful. The quality of the image is really bad.

Yeah, please share. That will be helpful.