SD3.5! Let's go?
SD3.5L seems to be the best of SDXL combined with (some) of the best of Flux. The results are looking promising! I am doing a test of artist styles right now and liking what I see. I think most artists present in SDXL are present in some capacity in SD3.5, and they certainly pop more (at least without other prompting)
Maybe it's time to rekindle my interest in expanding the artist set and doing some style clustering. What do you say? Want to make a workflow for collaboration on this? I think, and hope, SD3.5 is the last update for a while so it has time to cook like SDXL did.
Some examples. I am doing ~300 overnight tonight, will see how they look tomorrow after work.
These samples look promising! Style match isn't nearly as good as SDXL. But that doesn't matter if SD3.5 can maintain a certain style for various prompts.
I saw a video about tweaking the weights of certain network blocks to emphasize style. Now I can't find it. I haven't played with SD3.5 myself yet. Waiting for Forge support
You probably saw this:
https://www.youtube.com/watch?v=en-GMBIa-N8
or maybe this
https://www.youtube.com/watch?v=0ChoeLHZ48M&t=118s
Matteo is the greatest when it comes to real, open research into this stuff.
I agree that it isn't pure style copying with SD3.5. It's more like midjourney's "style-inspired" approach, which I think is more appropriate anyway. I think if you really wanted to copy an artist's style to the color, you can reinforce that with prompting. But that's not what most people want- they want some consistent styles with short triggers (names) that they can remix and play with.
That's why clustering will be useful. Say I like artist X style but I want to tweak it. It will be nice to have artists Y and Z (whom I have never heard of) that are similar but with their own nuances so I can play with all those aspects together.
I am about 850 artists into this series, thousands to go lol...