I have been using DALL-E 3 for my research, particularly its inpainting feature. However, since last week, I have encountered an issue when generating new images with inpainting. Each time I attempt to inpaint a small area, the newly generated image progressively darkens with each edit. I’m unsure whether this is a bug in the system, as this problem never occurred before. Notably, I do not specify any changes to color or brightness in my prompts.
Kindly let me know if this a recurring issue or will be fixed in coming future. Thank you for your time and consideration. I look forward to your response and hope for improvements to restore the exceptional quality that DALL-E is known for.
Here are the samples. The prompt that I have been constantly using is ‘Reconstruct the missing vase of an 18th-century Mughal-style fresco. Restore the broken side of the ornamental design to seamlessly match with the original while preserving the aged, cracked texture and authentic style of the fresco.’
You can see the progression in the image getting darker and unsaturated from the attached image.
Hmm, I’m not sure if I understood it correctly. When I look at the images, I can’t really see that they’re getting darker. What I can see is that in some segments, the vase is incomplete. This is most likely due to the limitations of the current stable diffusion system. It’s not perfect, makes mistakes in small details, and since randomness is part of the process, it can contribute to errors as well.
Since it’s a repetitive pattern, it might be possible to pick the best segment and multiply it with a manual editing software. Or, you could create a large image and then use that to multiply such a structure.
So as much i can say, it is probably not a bug, but simply the limitations of the current system. Some careful prompting and experiments could give you some limited benefit, but i think it is the current model limits. the smaller the details, the more distortions and errors will be produced.
Hi. I have been trying to use inpainting to reconstruct the missing section of the vase. SO I have to redo the inpainting over and over again to get the right output. In process, the images get slightly darker each time and by the time I have done 10-15 inpaints, I can see a clear result that the image got significantly darker. See the attached image. This was the result after multiple inpaints and you can clearly see that image is darker for some reason.
This error was not happening before. But for last two or so weeks this is a repetitive pattern. I tried creating another account just to check if its the issue with the account, but it still had this error.
I got a similar problem using Forge UI but found a solution: using JPG instead of PNG as the input image format. I use it as output too so I can inpaint over the generated image, and the darkening is gone.
I know it doesn’t make any sense but worked for me, hope it helps you on your workflow.
In so many processing’s, it could have various reasons. One of them is that a lossy format is being used, and with each new generation based on the previous one, information is lost.
It’s better not to perform so many iterations on a single image, but rather try to correct as much as possible, in one go using better prompting.
