10Noob v1
Remade with NAI models and better quality flavor models. Probably should've started there. They are tricky to merge, but all the functionality is still there. All the NAI debauchery carried through the styling changes intact; anthro stuff, artists, characters, etc.
Same usability: Euler/EulerA. The DPM++2M/SGM Uniform combo with 5-10 CFG at 25+ steps seems to be the golden zone. Full hires fix pass 0.3-0.5x with matching steps at 0.6-0.7 denoise adds a lot of detail without changing a lot. Lighter CFG can give softer but washed out looks, but also useful for hazy look and film grain. Use negative prompts with caution they can hurt more than help sometimes. Can also be used with DMD2 Lora and LCM sampler to shortcut to more realistic portraits using bigLove realism tags like real life.
Very heavy 3D or detailed Illustration look by default. The 'realistic' and 'photo' tags are still completely hijacked. But overall extremely prompt and Lora responsive. Still has bad lowres details and tendency to over-pattern sometimes but composition is solid and some of that can be used creatively.
1ll v2.0
Reworked from my cunetta mix. This is also a meth merge of essentially all XL model archetypes because why not. But, very heavily Illustrious leaning, more towards the Hassaku and Noob models. Everything except for latent anime characters and some artist styles, although some seem reachable.
Comfy - use Coyotte's workflow. Has all the needed refiners: https://files.catbox.moe/147f5i.json
DO NOT USE realistic
If you do that you're narrowing the tokens for no reason to the small 3D data batches. The model already seeks to be semi-real by default unless you prompt anime, linework, etc.
Push the real look even further until desired with excessive negatives. Results will vary, this is an extremely prompt intensive one. Same usability and weakness as cunetta. Hands and Face take many steps to diffuse even though you can get a clear composition out in 10-15 steps. This is an inpainters tool for sure. Other key weakness would be sex setups with multiple figures, but there's definitely working prompts somewhere.
CFG 2.5-9
Samplers like sa_solver in comfy seems to be decent only for upscaling detail. EulerA is still the only real solver for consistent details I've found.
This model is jailbroken and the weights are technically cooked, but not in a bad way. Still seems user friendly, but some prompt explorations definitely go sideways. A bit of contamination from 10XLP and lustXL models but prompt response seems intact. Score tags seem helpful sometimes with Pony Loras and Pony-like images/concepts. This was supposed to be a merge tool model for adding IL to realistic models, but it worked out on it's own. This ones an endless rabbit hole of possibilities, you can pretty much use any XL Lora on the site. There's way too many dynamics going on with this model for me to test so I'll unleash it for the slop. I am terrible with Il prompting so the preview images are my half assed attempt.
v1 - More scuffed Pony-like version.