Updated: Jul 17, 2025
base model
Version 2.0 Information:
A repack of the original Something Noob merge checkpoints without including Noob followed up by a merger with Hyphoria. Should be a more pristine Illustrious experience geared more to linework and less realism. It feels to be in a better spot for prompt adherence and flexibility. It works best between 3 and 5 CFG and sort of hits a style ceiling at 6, but you can keep going for more/different details.
Positives: masterpiece, cinematic lighting, depth of field,
Negatives: <only as needed>
I'm not doing the phi thing anymore. I started seeing it everywhere after I started doing that, and now there's always a guy smoking a cigarette outside my house. I know he's not real, but he says he is.
Version 1.0/1.1 Information:
This is more of an upgrade to Something Borrowed that was my first checkpoint merge. I've really liked the results from Something Noob but wanted to bring in some more realism, and this one delivers really well.
Thanks again to the folks behind the models used to merge Something Noob, and thanks to the following folks and their models that contributed to this merge:
The prompt adherence on this checkpoint is a bit insane. Forget to remove a keyword and your characters will start playing twister to get it in the shot. Seems to be a bit shy with nudity sometimes, but ymmv. I was struggling with this until I switched to Euler and then it's been smooth sailing. It starts responding well at CFG 2 and works well as you turn it up. It starts getting posterboardy at about 7 or 8. As always, FaceDetailer is your friend.
My personal settings have gone full nerd mode. I use 2.718 CFG (e), 1.618x upscaling (φ), 0.618 denoising (1/φ) for bigger areas, and 0.368 denoising (1/e) for polishing. Sometimes 0.42 because Douglas Adams. Also using Fibonacci stepping trying to stick to 8, 13, 21, 34, and (rarely) 55. It's stupid, but it's my current flex.
The only issues with loras that I've had with this one so far is that I don't really need to use them. Style loras get the job done, but the prompt does a ton of work.
Adding style negative seems to just make things worse for me. I've been getting better results getting "mean of empty slice" errors from an empty negative box.
One issue I need to look into is a center offset at generation. Subjects get shifted a little to the right in portrait and down in landscape at flat camera angles. Not a huge issue, but could be a problem for character portrait loras unless there are keywords I'm unaware of that fix the problem.