Training Experiment with a Single Image

If you're interested: One Image Is All You Need | Civitai

This model was not trained on real imagery of a living, or deceased, person, or depicts a character portrayed by a real-life actor or actress.

Image used in training:

Experience with PuliD

I couldn't get any satisfying results with PuliD, no matter how much tweaking I did. So, I decided to train with just one image to see if I could get decent results.

Results & Observations

I'm actually very satisfied with the results. Of course, there are some differences compared to LoRAs trained on multiple images:

• It doesn’t work as well with overtrained LoRAs.

• Most likely more stiff with facial expressions

Flexibility Settings

Feb 26: Actually just set the strength to what you think is right, it depends on what you're doing. You can go up to 1.

• For more flexibility: Set strength to 0.7 and use the trigger word "ry4n".

• For more likeness: Set strength to 0.85 and use it without "ry4n".

CFG Settings

• Without any LoRAs: Set CFG to 2.7-2.9.

• When using this with other LoRAs: Set CFG to 2.7-3.3. I usually keep it near 3. Really depends on what im doing.

Trigger Word Considerations

• I include "light gray eyes" in the trigger, but sometimes it’s better not to use it because it can push the image too much.

• Example: When using this LoRA with an ahegao LoRA, I don’t recommend including "light gray eyes" in the prompt.

• Note: "Light gray eyes" wasn't actually a caption in the dataset—it’s just something that helps.