DoRA trained on 364 works of pandain95, also on pixiv.
Trigger words are pa-ndain95 (required), gigantic breasts (optional) and pregnant (optional). Add white pupils and use adetailer for better eyes. Did not train on any futa works of this artist as it's a pain in the ass due to censorship, and not my taste as well. I'm sure you can find tons of futa loras if that's your thing.
First successful attempt at training a DoRA with frozen text encoders, thanks to @throwawayjm which I got the idea from.
Description
LyCORIS/LoCon module with preset unet-transformer-only, which freezes text encoders
Dimension 16, alpha 8, no convolution layers.
Prodigy optimizer with arguments "decouple=True weight_decay=0.01 betas=(0.9,0.99) d_coef=2 use_bias_correction=True safeguard_warmup=True"
min_snr_gamma 2
soft min snr (see attached zip file)
ip_noise_gamma 0.02
DoRA weight decomposition
No dropouts of any kind
scale_weight_norm 1.0
40 epochs, 3 repeats, cosine scheduler
--network_train_unet_only
batch size 64
flip aug
1024x1024
multires iterations 6, discount 0.3
Tagged with wd-eva02-large-tagger-v3
FAQ
Comments (4)
Illustrious Ver Plz
Yea, frozen TE works pretty well for simple concepts on SDXL models. It does take longer to train, but I'm fine with that. Trying a 600k image dataset with no TE next, but well see how that goes.
May I ask which lycoris preset do you use? "Full" preset enables text encoder modules, while "unet-transformer-only" disables TEs but also disables convolution layers. Also, how do I enable network dropout when training a DoRA? Lycoris reports network dropout is ignored while DoRA is chosen.
--network_module=lycoris.kohya
--network_dim=$network_dim
--network_alpha=$alpha
--network_args "conv_dim=${conv_dim}" "conv_alpha=${conv_alpha}" "algo=locon" "dora_wd=True"
Is what I use for DoRA LoCon. So any other options would be default values. It probably is ignoring network dropout. Haven't had time to see if I could hack it in.
Probably "Full", but I think that might be a GUI option only. If you train with frozen TE then it will ignore TE in DoRA. I checked and my DoRAs dont have any TE layers inside.