With massive credit to a mysterious wizard who was able to reduce the file size by 97%.
Both versions are basically identical.
Style model trained for use with animagineXLv3 https://civarchive.com/models/260267/animagine-xl-v3
For positive prompts I used: anime screencap, visual novel, (artwork), (masterpiece), (amazing work), (masterpiece), <lora:ke-ta_3-000003:0.6>
For negatives I used: (low quality, worst quality:1.3), blush, head out of frame, nsfw, nude, breasts, text, signature, nipples, cleavage
Description
FAQ
Comments (5)
nice work！but can you trim this model to 300-200M？it almost use all my vram. thx
If you can run it you should also be able to trim it, instead of potentially waiting days or forever. Supermerger with A1111 - I think - has a prune option. There's a couple different extensions that have prune options.
I tried pruning it with model converter but that only saved 13kb, not sure how to do it 😖.
@GenaBrown You should use Kohya_ss for that. Use the option Lora>Tools>Resize. Usually the default settings are good enough, but you can also adjust the dim/rank to whatever you prefer.
@hhks I originally downloaded kohya to try and train a LoRa but I still haven't gotten it to work. Resizing a LoRa is the first time it actually did something, and had no clue what I was doing either.