Almost entirely NSFW Dataset except for the splash art and that's likely gonna reflect on the output if my example gens are to go by. It's up to you to gen SFW or NSFW, though I much prefer the latter.
Trained last year in July, but the outputs seem to look better the more recent the model used for generating is. I tried training more updated versions but somehow the first try worked the best, which is this one. Only had a measly 20 images to train it with so you might wrestle a bit with the results.
Trained on Fluffyrock Epoch 99, Terminal SNR Epoch 72.
Recommended weights (trust me):
The LORA itself works best at 0.8 weight, however some tags work best with certain weight ranges. The main activation tag, khazix, works best at 1.0-1.2. You can slightly emphasize green eyes to 1.1, while antennae can be 0.7-1.0. insect wings can be set to 0.5-1.0 or omitted entirely while brown fur around 0.5-0.8. The rest of the tags work as intended, but feel free to shift around the weights to what looks best. Also put jungle in the negatives at a weight of at least 1.3 to avoid jungles and leaves generating in the background if you don't them to generate.
Almost all example images are genned with Restart, the best sampler.
{
"ss_output_name": "khazix",
"ss_sd_model_name": "fluffyrock-576-704-832-960-1088-lion-low-lr-e99-terminal-snr-e72.safetensors",
"ss_network_module": "networks.lora",
"ss_optimizer": "bitsandbytes.optim.adamw.AdamW8bit",
"ss_lr_scheduler": "constant",
"ss_clip_skip": 2,
"ss_network_dim": 32,
"ss_network_alpha": 16,
"ss_epoch": 10,
"ss_num_epochs": 10,
"ss_steps": 4000,
"ss_max_train_steps": 4000,
"ss_learning_rate": 0.0005,
"ss_text_encoder_lr": 0.0001,
"ss_unet_lr": 0.0005,
"ss_noise_offset": "None",
"ss_min_snr_gamma": 5,
"ss_training_started_at": "2023-07-29T15:58:16.665Z",
"ss_training_finished_at": "2023-07-29T16:34:26.560Z",
"training_time": "0h 36m 9s",
"sshs_model_hash": "4d8eeb54576fa9c99964cb536219fdeebc02f47367aaac2e2066400170b6260e"
}