Type | |
Stats | 404 |
Reviews | (19) |
Published | May 26, 2023 |
Base Model | |
Training | Steps: 4,000 Epochs: 1 |
Trigger Words | aditisharma anushkasharma evelynsharma nehasharma tarasharma |
Hash | AutoV2 C1851D0308 |
This was more of an experiment to see if it is possible to train a LoRA with two or more separate characters, each invoked by a trigger word. Turns out it is possible, and the output is pretty good.
V1: Trained on the Realistic Vision 2.0 model, so that's what works best. CFG higher than 4.5 tends to deep fry the images, so suggested that it be kept below that. Steps = ~20. Weightage of 0.4-0.7.
Trigger words: evelynsharma, nehasharma
V2: This actually produces 5 different celebs, all with the last name sharma. This seems to work best with a clip skip of 2. This doesn't produce the same fidelity of images as V1, but was a fun experiment nonetheless.
5/28: Uploaded modified V2 model. Used regularization images in the training, which allows the LoRA to produce less baked images.
Trigger words: aditisharma, anushkasharma, evelynsharma, nehasharma, tarasharma