Trained on Moss Base.
Trained with long warm up, very high peak LR followed by a cosine-like decay. Wanted to see if inducing a high peak can bake in style deeper while maintaining a low rank (rank of this LoRa is 8) and keeping total step count low/training fast. Afterwards I deactivated some blocks i deemed unnecessary and increased weight of others that imo carry alot of the face style.
I adjusted weights while using the following settings
dpmpp_sde_gpu
16 steps
CFG 5
trigger: @murata range
Helpful tags: painterly, film grain
Helpful prompt when using Turbo: The artistic style is characteristic of Range Murata, with clean lines and a soft, painterly color palette.
I will probably retrain it with a conventional slow cooking. But Its a cool LoRa so I wanted to share it.
Added also the last epoche of the training session. It has better compositions and colors imo, but eyes are more likely to lack the Murata-style reflections. Please download both LoRa versions and try for yourself.