I have been a frequent user of DALL-E and have always appreciated its creative potential and output quality. However, recently, I have encountered significant issues that have impacted my experience and led me to consider alternative services.
Smudged or Blurred Outputs: Many of the generated images appear as if the colors are smudged or blurred, reducing their overall quality and usability.
Unwanted Blue Tint: A consistent blue hue or reflection appears on the right side of the images, which was not an issue in earlier versions of DALL-E.
These problems have been frustrating, especially since the initial versions of DALL-E provided exceptional results. My colleagues and I, who regularly use DALL-E for creative projects, are disappointed by these recurring issues and are actively exploring other options.
I would like to know if your team is aware of these problems and whether there are any plans to address them in the near future. It would be unfortunate to part ways with DALL-E, as we have greatly valued its potential.
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.
This is already under discussion.
The more voices that speak about their perceptions, the better.
If you want to test the model data in use, create a moon. If the moon looks ugly, the new training data is in use, which has reduced the quality.
I have a different but similar problem in dell-e 3.
Some time ago I realized that even though the images looked perfect when in the chat, when I download it and enlarge it a bit I saw many imperfections like parts missing and some of the smaller details have been misplaced etc. I have used a few other AI image generators and I saw the some of them had it too.
I hope this was helpful.
This is another issue again and is related to the so-called stable diffusion process. Essentially, it is a de-noising process with AI data. These systems generally have problems with details and topology. This can only be corrected with an improved method or higher resolution.
However, color shifts are dependent on training data. Low-quality files were included in the training, causing corresponding errors like color shifts, chroma aberration, over-sharpening, Noise, edge errors, etc.