<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"

胃之上奇嘉O + 鬼頭サケO + 信じO + ピジO

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 donenesssample images are made with DPM++ 3m SDE sampler, karras scheduler with custom value 2.0 rho, sigma_max 15.0, sigma_min checkpoint default100 steps, cfg 2.0

Files with SHA256: 9826a74479103fdbd7879727f2810dc419244e4c9fb87f2301f6d052ca3d81dc