Training:
Using advice to keep a very low Learning rate when training Chroma LoRa , at only 0.0001 LR using cosine with repeats scheduler , for 50 images x 16 repeats x 50 epocs = 40,000 steps
Dataset of 50 images - 30 are face portraits by Ilya Kuvshinov and 20 are black and white manga faces by Tsutomo Nihei
Training images (50) , all are safe for work : imagebam dot com/view/GAG61J
Trained on Chroma HD
Prompt structure:
a portrait illustration by ilya kuvshinov <tags> <natural language prompt> <tags> a portrait illustration by ilya kuvshinov
For ilya kuvshinov style (30 training images) , and
a monochrome manga illustration by tsutomo nihei <tags> <natural language prompt> <tags> a monochrome manga illustration by tsutomo nihei
for tsutomo nihei aesthetic
//---//
The gray background in the LoRa training images have the same RGB as Tensor Art background.
Randomizers:
Exact training prompts , all are safe for work: perchance dot org/fusion-t2i-kuvshinov-lora-training-data-1
Randomizer: perchance dot org/fusion-t2i-kuvshinov-lora-training-data-2
//----//