Gemini flash image edit - degraded

Gemini API
, ,
1

Gemini flash image preview API - edit mode. We see a huge drop in quality and respecting prompt since flash image preview was released. Makes very often to much changes to the original image. Experimental model was/is really good compared to this. Spent a week trying to get consistent and good results.
@Google any updates on this? I read this in different groups online also.

Can’t go back to experimental, to small rate limit. If not google raises it.

2

Hi @Matt3,

Thanks for your valuable feedback. I already escalated image editing issue with our engineering team. Is there any issue with quality of the image you are facing??
If possible, please share some of screenshots so i can further tag your feedback in the current escalation.

Thanks.

3

Thanks for adressing! Can you give some feedback to what the plan is? updates on its way in very near future, higher rate limit on experimental etc.? The quality of generated image feels compressed more than experimental in some case but that we can work around thru upscaling. The main problem is that editing is not working sometimes and mostly if working it changes more things in image than experimental model. For example if want to furnish a room it could remove doors, change perspective etc. In experimental this worked . We tried many different prompts, simple and complex adding things step by step but not near the result we got in experimental. When we used experimental we only need one good prompt and there where no changes to room besides the furnitures added.

4

Someone wrote this on reddit which describes it well, before and after:

2.0 FLASH DOWNGRADE

BEFORE

  • consistently high quality image generation.
  • followed prompt more effectively.
  • added, changed, or removed details from the image effectively.
  • intelligent ai. had a good understanding of the users’ requests.
  • did not get confused from large prompts and did not omit many details if at all.
  • capable of providing more generations without sacrificing quality and/or detail.

AFTER

  • does not provide.. all of the above.

  • gives you one halfway decent image generation.

  • quality and detail of the image drops significantly, the more details you add after the initial prompt.

  • unrealistic image generation-- i.e. People with 3 legs or arms, mutated fingers or blurry faces, assymetrical and deformed body parts.

  • capable of editing uploaded images.

THE PROS DO NOT OUTWEIGH THE CONS. BRING BACK THE OLD GEMINI.