I'm not the guy who came up with this style! I just vacuum up catboxes and do a lot of tests/repros. I'm not at my main computer so I can't give you exact gen parameters, but the cliff notes version is:
- DPM++ 2M K or DPM++ SDE K w/ a high eta are the best samplers in general, the former was used here
- search seed space at whatever res is reasonable (512x704) before upscaling
- once you have a good seed hiresfix with ultrasharp or remarci @ 0.5-0.6 denoise (i prefer the former, ymmv)
- do a second diffusion pass in i2i on top of the hires, but with a simpler prompt focusing on high quality e.x. "gorgeous anime wallpaper, beautiful colors, impeccable lineart"
- for the second pass (this will be taking you from ~1k x 1.5k to 2.5k x 4k) you're gonna run out of VRAM if you're trying to diffuse traditionally so use the Tiled Diffusion addon with Tiled VAE enabled as well to conserve vram. DDIM seems to work best here though I've gotten good results with the two samplers above as well.
- using a style finetune helps a lot, Cardos Anime was used here
- when in doubt, search Civit.ai or huggingface, there are tons of great models/LoRAs and textual inversions out there and if you have anything specific in mind having some form of finetune helps a ton
Obviously you're going to need to know how to prompt as well which is definitely not something I can explain in a single post, just like any kind of art you just have to practice a bunch to gain an intuition for it.
P.S. I've recently started a patreon, if any of you'd like to support my work on this stuff. I'm a big believer in sharing all my knowledge, so most of it will come out for free eventually, but I gotta eat. [0]
[0] https://www.patreon.com/thot_experiment