SDXL 1.0 Prompt Guide | Stable Diffusion

Stable....I am a poor blind deaf beggar.

If you want more detailed image of portraits than go with analog film it makes face 8k and doesn't give you so much dot type of lighting in BG 👍🏻 and also if you want some specific details like glasses than consider it in the Capital Letters.

"photo of a womennnn". :o

Uneducated hindu person who doesn't even know how to use SDXL properly. Shame.

Highly unrealistic blue eyes, that would be the first clue these were A.I. generated images.

Thanks for that. I agree with your findings about the ratio. It took no long for me to get almonst perfect portraits.
But when it comes to full body shots, SDXL is still messing with hands, legs and even faces. Negative prompts from previous versions seem mostly to be ignored.

What really annoys me is that i could not figure out yet how play with the depth of field. It's fine for portraits but there are also cases where you don't want
this blur or at least want to reduce it. With a camera I'd just change the aperture from let's say f/20 to f/5

Btw what you sould try is img2img a SDXL picture with ICBINP model. You are flexible and get awesome results.
For Clipdrop free users: it also removes the watermark on most pictures. Did i write this?

Thank you for this guide. I'm not too impressed with the quality in 1.0. Like you and others have stated, the lack of facial detail and realistic skin pores/complexions is disappointing. But we also have to keep in mind this model is still so young. Look how bad SD 1.5 was just a year ago. This is where community development will take over with new loras and other customizations. My only gripe is the file sizes for loras in sdxl look like they're HUGE 800mb, so unless there's a way to further compress them, storage will become an issue for many.

I think maybe your sample size on the aspect ratio was too small, SDXL is trained on 1024x1024 so square images should be the best or at least the same. Also the quality is not "downgraded" the initial realese is not fine tuned for photo realistic images, it is a genreal model that can do cartoons and art etc. Wait for the community photo real finetunes if you want truley realistic.

Much appreciated. FYI, you misspelled "woman" as "womEn". You also wrote "...glasses and overcoat" instead of "glasses and AN overcoat". In my experience spelling and grammar makes a huge difference in output quality, especially when (as is the case here) the mistakes actually change the meaning of the prompt in a way that breaks logic.

Remember, these systems are all based on gradient descent, aka statistics. Bad spelling and grammar, and descriptions that don't make sense, are associated with bad training data, and that bad data is what will guide the output of such prompts...

Looking forward to seeing more from you, keep up the experiments! 👍🙏

I'm sorry but there is nothing remotely 'realistic' about these images. I think you are confusing 'realistic' with 'what is pleasing to me'. They look more like airbrush art than photographs.

But it is ignoring the prompt, I mean , you are asking for glasses, the image is wrong without glasses, right?