(AAA)(no longer Autism) (no longer Lucereon)

I will break your loras version:

It will break your ilustrious and noobai loras. Idk, maybe noobai ones will surpvive actually, have'nt tested. Illustrious loras are dead.

Upd: looks like simple loras are not affected strongly, and can be usable, thanks to people for letting me know. I was testig mostly style gloras, and they were outputting quite bad results(trained on illustrious). Hopefully characters work decently too.

This version is a marriage of horny BigG trained by @bluvoll and aesthetic Clip L trained by me. This time full training extent and Unet adaptation for it to even work(Otherwise image will be just noise). on top of AAA. Consider it a finetune.

This Unet and Clips are likely to ruin compatibility with components in merges, so merge carefuly.

Model is experimental, and has it's quirks, but also benefits:

Minor issues(think pupils) and ocasional hallucinations udner certain conditions.

Overfit to anime women, of course.

Excells at abstract prompts and, apparently, holding weapons? huh?

I had so much fun genning with it yesterday, ngl.

It is trainable, and it is likely that you'll need to retrain loras for it, if you'd like this base.

����̝â̶͑͠æ̸̋â̶äۃ̐�ö�۟���� version:

Base is NoobAI 1.0 now. Tencs were replaced with my own finetuned Clip L, and finetuned by Bluvoll Clip BigG, so only unet of NoobAI remains.

Early epochs of tencs were used, since further training requires unet alignment, which i don't have time, not hardware to do fully.

Params i use to generate:

Sampler: Euler A, schedule: Align Your Steps (modified to accept max sigma, used with Sigma = 29.2)

Steps: any you like, no secrets there. Good old default 20 will do.

CFG: 7 works good. Sometimes need 9, and can go higher. Experiment.

Update version:

Tencs trained separately. I trained CLIP L, Bluvoll trained CLIP G a bit. In most images i also used VAE i posted recently.

Portion of weights were tuned back, to improve compatibility with loras.

Im officially swapping over to Lucereon, made by my good friend @bluvoll.

Please check out his model page, since it contains list of artists he has finetuned into model. I'll even leave you a BIG link:

Lucereon Link

Lucereon Full - Well, as below, but much more training. (like 3.33x or smth)

Lucereon Early is very early version of training on top of it, but training progresses quite fast, so i decided to release early version for the time being, so you have time to get accustomed to it.
Goal is just 50k, since training is not full yet, and on-site free gens are set to 50 so you can experiment to your heart's content.

Lucereon is a finetune based on recently released Illustrious model. It has quite a few issues that are quite hard to overcome, since they are rooted in training method, and we can't match size of finetuning to fix them without someones' sponsorship, but whatever. Personally, i was becoming qutie sick of pony, especially of all the hardships rooted in design choices of it, which made proper training hard. Illustrious is trainable quite a bit easier, but it poses more issues to end user - you.

What are they?
1. Anatomy - Illustrious was trained with extremely high LR for a finetune, and it was baking text encoders, at high LR too. Due to dataset size, model was not outright killed, but final outcome is lacking in quality, unless cherrypicked. In particular - anatomy. Please stay on previous versinos of AAA, if you pefer stability. I will not remove re-start v2 from on-site.

2. Fine details - Due to hyperparameters and training practices, Illustrious is struggling with small details, particularly it fails to denoise wide shots, and instead creates quite apparent noise patterns. You can see them in my showcase images. This will not affect generic 1girl gens, but it is a big issue.

3. Stability - Due to blah-blah, you get the point. It is not as stable.

4. Lack of loras (so far) - Since it's a new base finetune, amount of loras that work with it are limited. Stay on re-start v2, if you want to keep using pony loras. I hope we soon will see more loras for Illustrious, or directly for Lucereon.

But there are also upsides:

1. No deliberate obfuscation, hashing, etc. - You can prompt things and not be scared that random 3 letters will trigger some random style or even character. This is also a big thing for trainers. Model seem to be trained as-is.

2. Less tag poisoning - Very important thing for me. Tags like abstract were deeply poisoned by data from furry and pony boorus, on which "abstract" mostly means "simple" or "plain". I need to explore more, but Illustrious was trained primarily on Danbooru data, tags of which i consider higher quality overall.

I hope with time we'll overcome issues with fine quality on it, because as a base - it is better.

Prompt structure:

Positive: Just prompt. Quality tags are not necessary.

Negative im using: (worst quality:1.4), (low quality:1.4), jpeg noise, compression artifacts, jpeg artifacts, scan artifacts, aliasing, english text, watermark

But this is just me.

CFG: 4.5-7

Sampler: Euler a with Align Your Steps schedule.

Description for older models below.

(AAA)Autism

Re-Start v2: base has been changed quite a bit, in particular it was switched from Autism to 3pxAutism by @bluvoll. Has also been slightly merged with Isolice for stability.

P.S. On-site gen will become available later for that version, when staff is up, please be patient. Faeia got it on, thanks :3

RS-lightning(4steps) - I don't like distilled models, but i was asked to make one for re-start. Here you go. I went with 4 steps one, as it was balance between quality and speed, but i recommend using 6 steps. Use CFG 2-3. Results will differ a lot, both in style and in quality from proper model. Please use baseline Re-Start for best quality.

Re-Start note: Trained from zero with adjusted params and a bit expanded dataset in one go. Took 148 hours :agony:
I really should stop training anything big on 4060ti.
This version should be more stable and perform better at standarcfg.

~~@novowels did not say that it's really good yet.
LET'S GO BOYS

v4 note: @novowels said it's really good. Thanks Novo

v3.5 note: Direct continuation of v3. Same amount of training done, but at finer values, and with better technique that im slowly moving from other trainer. Should fix some stuff, and be more compatibe with loras.

v3 note: different, less unstable training. Should have less issue than v2, or alternative v3(which i wont release, it's fucked, lol, and i deleted it). Creates bigger difference from ancestor, while being trained for less steps overall.

v2 note: it is experimental af. You might notice issues, and possibly worsening on some components. Or you might not. You might find it better. This is further merge of experimental training. Use v1, if you find it worse.

Recommended settings:
Euler a | Align Your Steps (Euler a AYS on Forge)
20-28 steps
CFG: 4.5-7
CFG for restart/restart v2: 4.5-6

P.S. all showcase images are done with just that, no adetailer is used. Hires fix performed with NomosDAT.

Finetune of PonyDiffusion v6 based on Anzhc's Anime Aesthetics dataset, and transferred to AutismMix.

Goal is to bring less 3D anime style through native means(i.e. direct finetune or something close to it), that lots of us are used to from 1.5 models derived from NAI, a bit closer, but keep Pony knowledge and loras compatibility intact, if possible.

Kofi - https://ko-fi.com/anzhc

This is mostly an experimentap piece that is used to lay down basic grounds for finetune of higher magnitude.

Brightness tagging (extremely dark, very dark, dark, bright, very bright, extremely bright) was used, but tags did not learn anything of use. Balanced iamges were not tagged.

Rest of details are yet to be disclosed.

Overall, i really wished for it to be more capable with abstract compositions.

Example of change:

License

Model provided for non-commercial use.

Whatever parts of Pony Fair AI license apply.

If you want to use this model for commecrial use, contact me in discord - Anzhc.