Type | |
Stats | 562 858 |
Reviews | (72) |
Published | Apr 30, 2023 |
Base Model | |
Training | Steps: 3,100 Epochs: 20 |
Trigger Words | aubreyheadspace |
Hash | AutoV2 6060A3BCDF |
The headspace version of Aubrey from Omori.
The result of my experiment with Lion optimizer. unet lr = 1e-5, text encoder lr = 5e-6, epochs = 20, weight decay = 0.01, betas = (0.95,0.98).
Results Summary
Perfectly usable at weight 1. Usable at weight range [0.9, 1]. Anything lower may diverge too much from the Lora.
For her usual appearance, use
prompt: aubreyheadspace, 1girl, shirt, short sleeves, hair bow, smile.
Stylized checkpoint models (Counterfeit, Pastel-Mix, etc.) seem to be hellbent on adding shorts/trousers on Aubrey. Idk if it's my dataset's fault for not including much pictures of her bottom-half view or whatever.
This could be helped by increasing the CFG scale to 10 and adding
prompt: bottomless
and/or
negative prompt: shorts, trousers
It may also help to replace the prompt "shirt" to "dress". (Is she even wearing a dress? Idk)
EasyNegative is recommended, quite compatible with other non-character Loras.
Though there's still (very) much room for improvement, the Lora itself is pretty decent.