Well I think you were the only one around here getting images like that in the first place
But, yeah, I too noticed that some of my older prompts are not working anymore.

I simply added "ariel view" in front of my prompt and began hardly getting any dogs, yet the angle in the images didn't look any different to me. So strange.

Is it just me or is Bing creating more and more errors, I seem to be getting more 3 legged ladies lol and arms that just seem to merge with a leg

I did not experience this so far, but got the impression that in general the quality of the creations seems to be pretty bad in terms of realism.

But that's one thing I constantly have with Bing, very inconsistent quality. Unlike Midjourney SDXL, where pretty much everything has the same level of realism when using the same model.

Deformities in Bing/Designer generally come from the complexity of the prompt in my experience.
If its a single subject sitting, kneeling or whatever, you are generally going to get good results however as you increase the complexity and include more than 1 interacting with each other, thats when Bing starts to get confused.
Ignoring filters for a minute, you can happily have a person pouring mess over another. however if you wanted 2 people kissing, hugging or insome other prompts I used to use, kneeling between legs, legs and body intertwined etc, then the model really starts to struggle. For some of my NSFW prompts I used to use I would have a high success rate 1 or 2 images generated but the amount of deformities would be 6/10 images.

In terms of realism, I think SDXL can be a lot better however SDXL really doesn't cope well with complex scenes. You can train LORAs to help and use controlnet for poses but ultimately you won't get anywhere near such complex scenes that is achievable in DALLE3. It really struggles still with interactions between subjects.

That's true and also DALL-E seems to have way better understanding of scenes, backgrounds etc. in general, that's why I mostly prefer it to SD/SDXL.

That's the big advantage of DALLE3 over other models, their underlying AI for prompt interpretation is well ahead of the game.
The other big advantage DALLE3 has is the quality of the image captioning.

This explains it quite well https://www.louisbouchard.ai/dalle-3/

But in essence, DALLE3's data set is absolutely massive. They scrapped the internet and other sources for images and then fed these into AI to caption. The captioning is miles ahead of other providers although others are catching up like midjourney.

This is also one of the problems with DALLE3. They image set is huge and includes a massive amount of NSFW material as there is massive amounts of porn on the web. There is also a massive amount of celebrity images in the underlying data set.

This is where I can see SDXL and the next models SD release improve on. LORA generation with well captured images is key to creating good models that we can customize.
The prompt interpretation needs to improve vastly in SDXL but combine the 2 and we will have a good foundation to build on.

Anyone else getting the "are you human check box" when you try to create a image this just started hours ago and it keeps saying error or failed

Yes, this is the captcha check they started rolling out last week. remember getting the white box. It was broken then but now working.
generally you will get it after a period of spamming prompts

Absolutely. In terms of quality and realism on the other hand, I see Midjourney ahead of all other competitors, especially with ther newest model. I can only imagine what a combination of these two minus NSFW filter could do...

I can see a future version of SDXL using a local Model for prompt interpretation as well. Some of the newer LLMs are ridiculous and approaching GPT4. At the minute you still need a 3090/4090 or a decent Mac with enough memory but those requirements are coming down.
A specifically tuned LLM would easily be able to run on lesser hardware for prompt interpretation.

Which models did you experience to be good at WAM and realism? I do not have a decent GPU but use hosted instances of SD/SDXL with a broad collection of models and LoRas available.

When a previously-working prompt string makes a picture or two on Bing and then gets the Dog over and over, I've found that sometimes going back to the last picture that worked and then hitting CREATE again works.

I've been noticing that too. Glad it wasn't just my hopeful thinking!

So with all the changes, what are people now finding are the best terms to generate that nice dripping coloured paint and the slime?

Can I just say: I love how you're talking technical stuff and as a byproduct just offhandedly create some totally stunning killer pics! Ok, now please do continue with your techy talk

Nothing much has changed regarding the filters and censorship however I did notice some improvement in terms of performance during peak hours. No longer do I have to rotate between accounts to utilize boost tokens.

That's just my observation though.

That's because because the image filter is so broken you can barely get anything interesting out

I got to experience the "different servers have different filters" thing this morning. I was working a couple of prompts around 1am, and they were doing great. Just signed in to that same account and every one of them got a hard prompt block. Awesome. So I made a couple of changes and they worked fine again. The problem? I named my characters to differentiate their appearance and pose. Names like "gwen" and "elle" to not take up too many characters. Once I changed the names to "woman" it worked fine again. So I signed out, closed my browser, opened another browser, and signed in to a different account. The exact same prompts with names are working just fine.

Ooh, now I'm getting pop-up ads! "Create faster with Microsoft Rewards". I can only assume that's the first sign of a paid model coming down the road.

They are already selling it under Copilot Pro for 20 USD monthly which gives you 100 boosts daily. Sadly with the current state of Bing Image Creator, I doubt many of us would subscribe.

I don't even see the point of boosts tbh. Yeah, my images load a little faster, but not significantly faster. And even with boosts, I still get days with extreme lag, or the "we can't generate images right now" message. The only thing I pay attention to boosts for is to know if my account has been reset after I've capped out on creations lol.

Well... it appears that Dall-E 3 got nerfed at some point again today. Very tame prompts that worked this morning are giving me the hard prompt block. I got a couple of them to work, but if I change ANYTHING, like time of day or lighting or camera angle, it also gets a hard prompt block. I just got my first suspension for prompts that were pulling 4 images at a time this morning. I can also say that running a working prompt after getting a hard prompt block does not save you from being suspended. It was a fun few months, but I don't feel like its really worth the minimal effort anymore.