Type | |
Stats | 43 190 |
Reviews | (9) |
Published | Nov 23, 2024 |
Base Model | |
Training | Steps: 2,000 |
Usage Tips | Clip Skip: 2 |
Trigger Words | goumang |
Hash | AutoV2 6E876BC896 |
Some time has passed, and there is more available to train on. As such, the LoRA has been retrained on a larger dataset, with captioning changes. There is now a PDXL version, as well.
(Trained at 768px with 70 images for ~5600 steps using Prodigy optimizer on terminal SNR FluffyRock e103, at 1024px for ~2100 steps on PDXL.)
Activation keyword is goumang; Suggested LoRA weight is 1. Lower than 0.8 may lose cohesion.
Recommended positive keywords: halo, feather hands, prosthetic leg, avian feet, long eyelashes, red eyeliner, kimono, slit dress, ear piercing ("collar" is available, but not trained well.)
Recommended negative keywords: cleavage cutout (Due to the dataset and PDXL training, if artist LoRAs are not in use, I suggest adding "aliasing, pixel (artwork), pixel art, oekaki" to the negatives as well, for PDXL specifically.)