Vadiryi's OC版本v1.0 (ID: 520142)

This is a LoRA I trained using 64 images of myself, mostly selfies. It works really well on more realistic checkpoints but it can do a more cartoonish/anime style as well if you work the prompt a little.

It is highly recommended to use this primarily in inpaint to replace a characters face as the images it can generate in txt2img are fairly limited to selfie style portraits.

The training data included mostly pictures of my natural black hair, but also some of when I had my hair dyed blue and silver, so it can do those three colors fairly well

Recommended to use weights 0.4-0.7 for best results since it can overpower most prompts. When in-painting, I found that a denoising strength of 0.4-0.65 works best. And try to make sure the hair color is already pretty close to what you want in the final image since it doesn't seem to be able to change hair colors that well in in-painting.

Trigger words: vadiryi, white hair, blue hair

Last but not least, since all the training data is of real photos it really likes to make photorealistic results. I've found that it overdoes my features and makes contorted messes at higher weights. The best way I've found to combat this is putting "anime style" in the prompt and "photorealistic" in the negative, as well as using checkpoints that are tuned for more anime looks from the get go.

描述:

训练词语: vadiryi,blue hair,white hair,1boy