I'm not a fan of describing models, but this is my last agony while waiting for SD3, so I'll tell you about savsdxl-5.2.3
- Increased detail and sharpness of lines.
- Great light
- High dynamics.
- Decent nsfw.
Compliance with artistic requests and not quite the right adherence to the promt that you will not even notice, as they say a feature, not a bug, but judges you when you notice it at more than 75 tokens.
In order to avoid questions I will tell you at once: this is a mix of the previous model 4.0.5 between Xi and delta.
In order counting can realize that I created and tested two dozen versions before coming to it. No, I don't need donations, just try this model.
Compare it for yourself with one seed on your favorite queries and favorite sdxl models, maybe you will like its spirit as I liked it!
From experience: 100 steps can give you a good result and just as well you can get a good result in 30 steps. Start your comparison with small.
Please let me know if you have any problems with the samplers, because I didn't have time to test all of them, use mine for starters.
And a secret for all who merge checkpoint: I was not aiming for a certain result, I expected only beauty at the end of generation. In alpha and beta it was RANDOM (R), but when choosing models for merge I expected to get a good result. Choose good models for merge.
Description
FAQ
Comments (6)
It seems like the majority of XL models are leaning towards artistic/semi-real/fantasy. XL in general seems to be able to do anything that's not photo real with ease. The real standouts are going to be the ones that can do photo quality in creative and flexible ways as well. I will keep an eye on this one, but it seems like I have this style covered with a handful of others already.
It's still very new. This time next year it's going to be wild (it won't take that long).
@sgqw0n8ib243 Yes as people train in new data it will be amazing. There are already advances coming from left field that are blowing my mind, like the Perlin Noise Generator for ComfyUI. I have been getting hi res photo quality out of just about any model with that, especially portraits. I'm happy to see good LORAs and working controlnet models as well, it is all coming along quite nicely.
This model can do regular photos as well. Have you tried it at all?
@WTFusion I'm a bit behind on testing models, I just wrapped up testing and categorizing my latest 1.5 acquirements, and am starting with XL now. Problem is, even many of these early models are blowing me away, and I just get wrapped up in a particular one trying different things. CrystalClear was an early standout, and it still amazes me. Then there was RealitiesEdge, Nightvision, Juggernaut, now a new Acorn release and a bunch more. A model that *can* do photorealism is fine, but even in the 12 days since my original comment, the whole landscape has changed. At this point, I think flexibility will be the key, how a model adapts to different styles and moods, and creates things at the edge of our imaginations in breathtaking detail.
People think AI is going to kill art, but putting art in the hands of everyday people is going to elevate art and culture into the stratosphere, and drag some elements of humanity down into the depths of degenerate sludge as well. In the coming generations, we will know the honest truth of humanity, whether we become truly great, or fail as a society. I appreciate what these model developers are doing regardless of some criticisms, because all of this is moving us towards our ultimate potential. Or some shit like that, maybe I'm just high :P
but not now :D there are a lot of realistic models now