Is it normal to want to augment the data?

[TomLucidor]

First time LoRA tester here, I got a weird question:

1. Is it normal to want to augment the images (e.g. randomly tilt by <10 degrees) to increase the usability of LoRAs?
2. Would cropping for focus (e.g. clothing, poses) be useful to train more comprehensive LoRAs? If so, how can that be automated?
3. Would color correction or other filters (e.g. contrast, brightness) improve LoRA training?
4. How do you deal with the asymmetry where there are few back-facing images?

[O-J1]

I don’t have the time to answer these 1 by 1 but I will say this; augmenting is a last resort, a valid one but still a last resort.

[efhosci]

I've experimented with some of the image augmentations, I can't say for certain how beneficial most of them are but they don't seem to hurt the results as long as they aren't overwhelmingly strong. Usually I enable flip+jitter, and have random brightness/contrast/saturation at around 0.1-0.2. I did notice that the "random rotation" augmentation caused it to learn black corners on the images so I've avoided that (not sure what the difference is to "random rotate and crop"). Hue can more easily cause a mismatch between the image/captions if for example it turns a "green" shirt into a "blue" shirt so I usually avoid that or maybe keep it ~0.05.

Ultimately if the image with augmentations still looks acceptable as the subject, and the captions still accurately describe the image, it should be fine to include. Photos will naturally have variations in exposure/color balance/etc so similar variations are ok, for animated/cartoon images it may need to stick more strictly to the source palette if you want to recreate the style more accurately at inference.

[TomLucidor]

• "random rotation" = tilted camera
• brightness + contrast + saturation = lighting + exposure
• "hue change" = color temperature + color tint (rather then color wheel rotation)

Maybe there is a way to make color change to be along the two axes of color... but still kinda surprised how diffusion models are not rotation-agnostic.