Introduction

This is a small-scale aesthetics fine-tune, based on subjective and quite arbitrary choice of images done by me.

This model is meant to represent rough estimation of what aesthetic i like(not necesserily 2.5d), which is, admittedly, not something out of this world, but possibly could provide you with something new.

If you wish to support me - https://ko-fi.com/anzhc

(probably) Approved by Cabal

Additional License Statements

Checkpoint is provided strictly for non-commercial, private usage.
I do not allow third-party platforms to host, reprint, distribute and otherwise utilize this checkpoint. (Such as, but not limited to: Pixai, Mage, Seaart, and others)

Basic Checkpoint Information

This is a standard anime training, so you will be familiar with how to use it, if you have experience with almost any other anime checkpoint, but there are couple differences that might, or might not, benefit your way of generating.

Positive tags - They were not used in training, so common tags like masterpiece, best quality would have less effect, and are not required to receive awesome images, as you can see in showcase gallery.
And tags like award winning, photorealistic, 4k, 8k should be banned from anime checkpoints, i swear...

Negative tags - Same story. They were not trained, so you can use your default, something like (worst quality, low quality:1.4) as golden standard of negative proompting.
Please don't use negative embeddings. Or at the very least don't use more than 1.

Resolution - This checkpoint was trained in 768 bucketed resolution, which does mean it is generally more proficient in higher reso than usual, and generally it does work with base reso like 768x1152, but keep in mind that this is a small-scale training, and does not target majority of model concepts, so your experience with 768 base resolution will vary from prompt to prompt.
I would suggest to keep it 640. (i.e. 640x960)
And yes, it can help loras trained in 512 to be higher resolution, but be realistic with expectations.
1 million steps, while a substantial amount, is not enough to fundamentally change model in all aspects.

Artists, characters and other - No artist, character and metadata tags were left in dataset, so you will not be able to emulate any particular style, hopefully.

Data for nerds and other cool people

Dataset

~10600 images, that i manually picked. No mass scraping was utilized.
All Artists, characters and metadata tags were pruned from tags.

Tagging was filtered through my filtering script.

Training

Trained for 100 epochs in 768 resolution.

Trained with usage of REX scheduler, decoupled weight decay for unet and tenc(set manually) and modified Offset Noise approach, which marched value from 0 and used randomized multiplier each step with conservative bounds.

Normalization layers were frozen during training.

Clip Skip - 2

Took 130 hours of non-stop GPU torture.

Special Thanks

To @FallenIncursio who is pretty much sole reason i have hardware to train dreambooth and datasets of that size. Thanks man.

To @novowels for being cabal leader. (and having no balls :clueless:)

To @PotatCat for being hyped about this checkpoint for last week.

And to all other caballers who did not participate in this checkpoint creation directly, but were providing experience and support throughout second part of this year:
@Manityro, @Shippy, @EDG, @ChameleonAI, @wrench1815, @dttiger, @yomama123556778, @richyrich515, @zetsubousensei, @justTNP, @bombasticmori, @za4beqsbv36z2s889, @AstreaPixie
Love you guys. Happy new year. Tomorrow.

P.S. Modern AAA games are shit, better spend your time making waifus on AAA checkpoint

:clueless: