<a target="_blank" rel="ugc" href="https://danbooru.donmai.us/posts?tags=inoue_kiyoshirou">inoue kiyoshirou</a><a target="_blank" rel="ugc" href="https://danbooru.donmai.us/posts?tags=kitou_sakeru">kitou sakeru</a><a target="_blank" rel="ugc" href="https://danbooru.donmai.us/posts?tags=shinjiro">shinjiro</a><a target="_blank" rel="ugc" href="https://danbooru.donmai.us/posts?tags=pija_%28pianiishimo%29">pija \(pianiishimo\)</a>dataset composition with a ratio about the same as equal weight128/128 network and convolution dimension/alpha5e-5 u-net and 1e-5 text encoder learning ratepolynomial learning rate scheduling with a power of 0.5learning rate cycle reset every 500 stepsno min snr gammabatch size of 1nesterov 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"

Files with SHA256: bca6cd5e09b03f4fcb950e76033c03e870cf7ef5f41ead3f3ad7752a1175e289