dataset composition with a ratio about the same as equal weight
128/128 network and convolution dimension/alpha
5e-5 u-net and 1e-5 text encoder learning rate
polynomial learning rate scheduling with a power of 0.5
learning rate cycle reset every 500 steps
no min snr gamma
batch size of 1
nesterov adamw optimizer (torch.optim.nadam) with args of "betas=[0.9,0.99] eps=1e-08 weight_decay=0.1 momentum_decay=0.004 decoupled_weight_decay=True"
Description
notice consider the substance aspect of the dataset, from now on me personally will not suggest and recommend using the trigger word(t1r1i1g1g1e1r) for it, since the dataset is made up by mostly comic book pages with multiple panels and multiple characters, this is not what you want in your image. me personally did see dramatic quality difference using and not using trigger word, and I suggest you do the same.
.
4 loras different doneness
sample images are made with DPM++ 3m SDE sampler, karras scheduler with custom value 2.0 rho, sigma_max 15.0, sigma_min checkpoint default
100 steps, cfg 2.0