This change both affects the API and AI studio. It’s shown that the last update to this model was in May 2025, and since then the model has been performing notably worse in the sense that it’s not nearly as coherent as what I had seen from it before this apparent update, firstly I knew there was a change made to the model when it would automatically change my prompt starting each response with “I will generate (it’s version of prompt)” unless I tell it not to, this doesn’t get solved with the advanced settings either, I’m sure in the background the image model does make enhancements but the language part of the model responding with revising my prompt doesn’t benefit the output, before it would only include the image in the response, and the generations which these previous parameters appeared much better, and much more accurate, on the topic of accuracy, it’s capabilities with image prompts have been widely reduced as well, I used to be able to give it multiple image references and it would use all of these reference to create one unique image, now more often that not, If I send it more than one image and tell it to edit virtually anything it not only fails to perform at this task but does nothing but warp one the various existing images I’ve gave it, even if I tell to only use these as reference or to just remove this one from an image, the image model doesn’t comprehend, and I’m very certain this is a knowledge reduction though unsure if it was on purpose or not. Even if it was on purpose I do strongly suggest to revert the model to it’s version previously before the May 2025 update as it was previously ahead of all other models at prompt adherence. I’m sure it wouldn’t have become so popular if it was performing like this current version when it initially went public. There is a solution to at least let these versions run simultaneously because they behave far too differently to be passed off as the same exact model, though I don’t see a purpose this current version can serve if the previous version was better overall.
Faceing same issue “Modified by moderator”
Yaah its true. When google launched Gemini 2.0 flash native image generation model. Its like hot cake until the goat image generation by OpenAI. in recent update its seems like google is lacking in it. I ensure they will fix in I/O.
I agree, the previous version is muuuuch better. in 1/10 it edits the image and keep the original image. I cant use it anymore. Revert or fix.
I should add that I’m requesting for the model to be reverted, after some consideration, “fixing” a nerfed version doesn’t make any sense than improving upon the one that performed well. I’m also adding to my request to revert the watermark back to being the gemini logo instead of what is currently shown, I am also adding to say this updated version is also running on gemini.google.com.
Furthermore, I don’t even know if the May update is when the reduction in quality was made, it could been sometime in late April, I remember the model was performing at it’s peak in late March for reference.
Any updates to this in I/O ? i cant see anything regarding image edits. Not in imagen either
Hi @all,
Thanks for the detailed feedback on the image generation quality issues. I will make sure to bring this to the team’s attention.
Could you please come back some plan or info for us.
Is there a way to workarund it ? different prompts, API settings or If you can higher Ratelimit on experimental to its solved? Cause now with no info we must look at other providers /solutions
Thank you. Also the version of the image editing model that’s being used by gemini 2.5 flash on the gemini website currently is the best model google has ever put out, it just lacks polish but it has far more potential than imagen.
I tried many different prompts and ended up asking model itself. Now it says it cant edit uploaded images.
Hey just a quick update but the version on gemini.google.com also seems be an altered version, Though it’s still a lot better than most recent version I don’t think anyone including myself minds it as complaints started happening after the current version dropped, the launch version of the model in march was the best performing every update seems to progressively make it worse.