Typ
LORA
Veröffentlichungsdatum
2023-07-17
Basismodelle
Other
Versionsvorstellung
Back in the day (like 6 months ago), we only had meme science when it came to LoRA's. This is a more sensible training of the style at x512 resolution with aspect ratio bucketing on top of the NovelAI model at clip skip 2 without a vae enabled. It seems to listen better now and it's much smaller (and could potentially still be effective at an even smaller size, but I'll leave that up to you to explore). Trained on an A100 with batch size 10 using DAdaptAdam.
