(W.I.P) New LoRA - Hoshino Ruri (Martian Successor Nadesico: The Motion Picutre – Prince of Darkness)
Added 2023-07-01 20:36:29 +0000 UTCHi. I'm posting my tests with my new LoRA. I'm trying to find the correct epoch to publish.
Would you like me to post more Work In Progress? (This is a test post to see your interests).
My training settings are:
- Base Model: NAI
- 10 epochs
- 4100 steps (410 each epoch)
- 41 images (repeat is set to 10)
- DIM network: 128
- Alpha network: 16
- Learning rate: 0,0001
- Txt encoder learning rate: 0,00005
- Unet learning rate: 0,0001
- LR Scheduler: constant
- Optimizer: AdamW8bit
- Precision: bf16
- Max Resolution: 576x576
Steps formula = repeat * images * epoch
I'm not using batch size to increase speed because it leads to overcooking sometimes.
I've used WD14 captioning with General Threshold 0,2 and Character Threshold 0,4
I personally prune a lot of tags and adjust them based on my knowledge on what is important to retain a style and keep characteristics without overflowing with useless tags.
Example: lots of blue hair, grey hair, white hair, purple hair... All became purple hair.
The same applies to eye color. Also I did some corrections to style tags. It doesn't need to know it's 1990s_(style) or retro_artstyle. There is some objects too that needs better description. Also selecting the images is based on experience.
A tutorial for style might come soon when I train with specific style.
