Impasto/Finger-painting style with visible painting strokes.

Notes:

Experimental LoRA that fed AI generations back in to augment the dataset (turns out there aren't that many anime-style pictures in impasto style!). If that sort of thing interests you, you can replicate the benefit by:

• Train the target style on your small and curated dataset.

• Create a small collection of varied prompts with the content you want (in this case, anime characters/style).

• Apply LoRA scheduling to strongly apply the LoRA during specific steps (in this case, I raised the strength sharply during whichever steps forced it over the content. In my case, I used this loractl schedule in NeoForge: [[email protected], [email protected], [email protected], [email protected], [email protected]]. By using 0 strength during other crucial steps, this was my desired balance for retaining and representing anime style. You can adjust/scale as required.

• Take the "best of" gens and use them to augment the dataset (you get captions for free!).

The first run was ~660 steps, the second run which included AI generations converged really quickly in 120 steps, same settings in both runs.

This indicates to me that there is value in carefully and thoughtfully widening the distribution of a dataset to target more specifically the things you want to eventually gen.