Some reorganizing
This will now be the place with all the synthetic datasets and associated loras.
(I'm not gonna delete this, but let's forget this happened, go over to Sloppy Moments and grab a lora merge for more or less instyle output)
There was an attempt!
I'll lay it real for you, guys.
Anima ALWAYS finetunes slightly more cartoony than training data.
You really need to amp the training data in realism to land on believable render style.
That out of the way...
I made a dataset from a recently discovered lora combo of realistic-style + beastman-style + hero-style.
Due to reason prefaced above, the dataset doesn't quite train to the same level of output mentioned lora combo does (aka, I still prefer the lora combo).
But since dataset is already made, no reason not to upload it!
Did a sample training as usual (10 epoch this time). And it looks a lot like 'toon style' but with beastman-style'ish muscle, more shading and (ugh) 'dramatic lighting'. Personally, not a fan.
Possibly, that happened because I used WD tagging and style got dispersed between random tags. I'll do another attempt but with no tagging and see if it yields anything better.
Until then, it is what it is - a cute slop picture set xD
A round of applause!
Billy: But, but-! Mr Proxy Sir, AI image data is more blurry than OC data!
Answer: Well, Billy, I train at 640x640 - all data is blurry at 640x640.
Synthetic Pixel Art!
200 image dataset of male-centric synthetic pixel art images and a sample 20 epoch lora (contains male nudity).
Images were generated in SD1.5, then later normalized with pixel fixer and upscaled back to original size.
Style is quite basic, but sample lora mixes well with other loras.
This was on todo list for a while, ever since Sloppy Moments pixel style sadness. And now it here!
Enjoy.
(funny how entire dataset 2 MB xD)
General Anime
General 300 image synthetic anime dataset and a 30 epoch lora trained on that dataset.
I was doing some random prompting (with a different lora) and ended up spawning a bunch of cool images of various anime characters. So, I made a dataset for yous.
This dataset will finetune Anima for general consistency.
It's male only and NSFW (as usual).
I trained it untagged, with just a trigger as a formality, but you can auto-caption it if you want.
Enjoy.
(I did wipe metadata again, but dataset prompts come from a free online deepbooru wildcard file, you can just google for that or ask an AI to find it)
'chartact' is new - male character with a mild hint of Akihiko Yoshida's exaggeration (FF Tactics).
Dataset is untagged. After a bunch of training, it almost seems like tagging barely matters for Anima. At least with such narrow datasets as these.
Trigger is unnecessary either. I add it "just in case".
'toon' has two datasets, one was made in SD1.5 and the other later in Anima from lora trained on first dataset.
I did not retrain SD1.5 dataset on Preview3. At least not yet. Busy with other things and not sure it's even necessary.