99% Synthetic data around comics, with a few small salts in of actual comics content to boost the style. not QUITE the style I wanted, that was more my fault since i varied the style of the content I THREW in. It's another more "SD VARIANT" style, but a good CHUNK of the base is newer MJ comics styles I threw together in the mashup bucket.
The Illustrious one has a frozen UNET and we're going to re-train the data asap, we used our lora trainer for the Illustrious one.
It DOES PUSH A LITTLE and it's not "BROKEN" but the metadata may not show resources as it's an A111 style output and my discord bot hates me.
Join us on this journey. Welcome to Ktiseos Nyx.
Our Discord Invite Link: https://discord.gg/HhBSvM9gBY
Earth & Dusk Media (Not AI) https://discord.gg/5t2kYxt7An
Our website (WIP): https://knx.neocities.org/
How to support:
Patreon: https://www.patreon.com/earthndusk
Membership: https://ko-fi.com/earthnicity
Membership / Ko-Fi: https://ko-fi.com/duskfallcrew/
Membership / Ko-Fi: https://ko-fi.com/OTNAngel/
Buy Me A Pizza/Coffee: https://www.buymeacoffee.com/duskfallxcrew
Merchandise: https://duskfallcrew-shop.fourthwall.com/
Description
Frozen Unet, made a mistake in the preset generation for our UI- it WORKS but it's waffley.
This was the mistaken issue with the Scheduler not quite likely being understood in our wrapper setup:
ss_output_nameSynomix-V2ss_sd_model_nameIllustrious-XL-v0.1.safetensorsss_network_modulenetworks.loracustom.batch_size3ss_gradient_accumulation_steps1custom.resolution1024x1024ss_optimizertransformers.optimization.Adafactor(relative_step=True)ss_lr_scheduleradafactor:2e-06ss_clip_skip1ss_network_dim32ss_network_alpha16ss_network_argsundefinedss_epoch10ss_num_epochs10ss_steps3480ss_max_train_steps3480ss_learning_rateNoness_text_encoder_lr0.00005ss_unet_lrNoness_noise_offset0.1ss_adaptive_noise_scaleNoness_min_snr_gamma5sshs_model_hash6e3be67a72ef16d211c696709b7f8c59c7e3f3c3ceb580baae136ce40c993d16custom.training_time2h 39m 16s