Not sure if this is affecting other AI image platforms but I recently tried to generate a few basic images through Bing and the results were significantly worse than they were from about a month ago.

These are just two of many, but you get the idea. These were very basic prompts and Bing still couldn't produce proper images after dozens of submissions and tweaks. The faces in particular look... over-rendered? I'm not sure how to describe it, but it's consistently bad.

I remember it being this bad over a year ago, quickly improving, and now it's suddenly just as bad as when it was first introduced.

Yes, it looks like bing has been updated to use a new lesser model of Dalle3. The consensus is it's using a distilled turbo model now which is faster but lower quality.
At this point now, I'd be looking at other tools.

https://www.reddit.com/r/bing/comments/1hhosme/bing_image_creator_now_uses_pr16_the_latest_dalle/

I have gotten nothing but Dogs for more than a day. All using previously successful prompts.

Yes, that is very cartoonish. Pixar might like it for their next release. I think there are going to be some massive lawsuits about plagiary and AI coming up so they are toning down the neural network recovery of real human image. People who might not like their face shown doing something my might not want. To be honest I prefer the real thing, but having said that the technology is very impressive.

T x

It's my first AI group post so don't shred me!

The previous post touches on this subject but I'd like to elucidate further:

AI has, if you've not already noticed, started getting a bad reputation. Some think it's just lazy, others think it's putting real artists out of work, others are angry that much of it (not all) has been developed by stealing artists' content and people's pictures without permission, others decry the amount of energy used to run the servers and computing power it demands. Others fear it getting impossible to tell between real images and fake generations that could be used for slander and political provocation. There is pushback against it both in formal legal proceedings and from user mass action. One form of the latter comes in the form of 'poison' apps that alter the image files in a way that causes AI to misidentify the visual elements in the image, without causing noticeable degradation to a human eye view.
The AI companies are already scrambling to either fend off lawsuits or else get licensing for the base images. That doesn't come cheap whilst it limits the pool of sources for generation. This is likely a big reason why you're seeing less variation in what you are asking for, not that the body types and appearance of AI females was ever spectacularly diverse to begin with.
It's also very likely that popular AI providers have already begun the process of enshitification we've seen with things like streaming, whereby AI is increasingly going to become a premium service but you will progressively get less and less for more of your money and possibly the acquisition of competitors to limit choice even further. They have likely calculated that a lot of AI traffic is going to be for, shall we say 'personal pleasure' and count on you keeping clicking and spending to try and get that elusive perfect image you desire, which will remain elusive because that's the business model.

I think there is a tendency to over analyse. There are a couple of factors at play but the most prominent is is the simple fact that AI can be exploited which opens up companies for litigation. When Dalle3 released, the quality was off the scale compared to what we see now and there was very little censorship. Within the first week, NSFW celebrity deep fakes were all over, gore and other propaganda were easily possible. OpenAI as a US company and MS who offered dalle3 through bing back peddled so quickly at the thought of litigation and introduced increasing layers of filtering.
The reduction in quality also followed as they tweaked these filtering layers. It's a lot easier to identify fake images when they look plastic.

Cost is also a huge factor, there is finite amount of processing available to OAI. With that they're training newer models such as o1, o3 and providing a services. Each model after release is later distilled to produce faster generation of responses but the quality takes a dip. Faster generation is a cost saving.
Its like with flux, I can use a non distilled model and an image takes 120s to generation, switch to a distilled model and it takes 30 seconds. Turn down the number of steps and it takes 12 seconds etc.

I'm not dismissing other factors but the core issues are economics and litigation. Incidentally, GPT-4o is capable of native image generation being multi modal. the outputs released and shown a major step up from Dalle3 but they've not turned it on. Instead, GPT-4o will revert to using dalle3 still when used.
The reason OAI gave is there are other bigger priorities.
Google's new ImagGen3 model is arguably better quality than dalle3 too but does carry censorship.
Grok2 is now using a new image model which looks to a be finetune of Flux. In some cases better than Dalle3, less censored in some areas, but not trained on enough data.

These were all pre filtering with Dale3 btw