These days DALL-E 3 on ChatGPT has been giving me results that are of lower quality than before. I’ve noticed increasing amount of smudges and artifacts, and some results looked more like bad Photoshop job. This has been going on for 2 months now and they have been completely silent about it.
Strangely, when I use DALL-E 3 through other sites (such as Copilot Designer and Anakin.ai) the quality comes out great and is the same as before the incident.
I’ve created comparison images to show you what I mean. Same prompt, I checked.
I had to create this comparison of four generations through two different platforms since I had the misfortune to encounter a moderator of the OpenAI Discord Server who dismissed this issue as because of seeding…
It seam it is not the same DallE system.
I detected the difference a while ago. and what i call the “template effect”.
You can find here discussions about the issues. Including the birdshit-moon, and the nonsensical light in complete dark scene. And the back glow is behind all.
If you want, create a complete dark scene, and a object in it, you will see a stupid light source in every scene, witch ruins all images. the change happens somewhere after 2024-06, i think.
I seams not all DallE System have it…
I’ve created two comparison images, with red circles added on the noticeable smudges and artifacts to highlight the problem I have with the current quality of DALL-E on ChatGPT
I recently had a very unpleasant argument on the OpenAI Discord server with two ignorant people who think DALL-E on API and ChatGPT are the exact same when there is a difference in quality. They even doubled down on their claim after being shown proof. I swear it’s this type of mentality that prevents this issue from being solved.
I have no experience with other systems besides ChatGPT and DALL-E accessed via browser. However, I speculate that different training data might be used, possibly even from different cultural environments (IPs / countries), under the same system. (It could be tested with pictures from culture specific background. for example, you test a prompt with typical Asian background in a other non-asian country, and check if you find “template effects” in Asian countries, not existing in other countries.)
The best proof for me that there are different data sets used is the ugly birdshit moon. In the browser version of dalle it is ALWAYS the same ugly moon, no matter in what picture it is a moon generated. With the moon you can proof it the easiest way, that the weights data are not the same.
I’ve also noticed the white spots. There may have been a nightshade infection at some point, and it’s possible that the white spots are remnants of that.
But at the moment, this is just speculation. As the developers generally do not answer questions. This policy of non-communication is not very helpful, so we are just left speculating.
