<h3 id="thank-you-mr-astromahdi-from-gpu.garden">thank you mr astromahdi from <a target="_blank" rel="ugc" href="http://gpu.garden">gpu.garden</a></h3>dataset images was from somebody else, I used this dataset as a test on how well Anima learn heavy painterly styles. It learn well.(Below is about Anima Preview 3)V1: Usual slightly overcooked setting ( adamw8bit kahan 2e-4, batch 8, 50 epoch, 16 dim/alpha, 100 image target, reduce block 0-7 to 1e-4 and dim 8), accurate to style but medium forgetting (like it cant really do pokeballs anymore)V2: Test setting (same as above but reduce block 0-8 lr to 5e-5, block 9-17 to 1e-4, block 18-27 to 2e-4), weaker style accuracy but less forgetting (can do pokeballs lol)But honestly I'm not satisfied with either scenario, but since I dont do regularization it's expected, you can try using V2 with 1.2-1.3 weight to get the style better without the worse knowledge forgetting (be aware it will become unstable still)DisclaimerI respect the artists who honed their crafts for years to achieve these styles among all else, if any artists whom LoRA i've made from shown displeasure/objection to their arts being used as AI training material or if they request a takedown notice on Danbooru , I'll delete the LoRA with no further notice.

Files with SHA256: f196f60036ec8aa1d3e01836687ee22912a09ca8bf48aa7acd59f41f34ca9efb