Dataset. Number of images

[tottem]

This might be a silly question, but if I want to use a configuration for 16GB of VRAM without modifying any parameters, how many images should I include in the dataset? I assume the default parameters are directly related to the number of images.

[sconnerart]

Not a pro here... but in my experience, the number of images doesn't have much or any impact on VRAM usage. Batch size does. The number of images in the pool just means you have to take more bites out of the data to finish your meal. The epoch being the meal here... one full pass through the dataset.

I have trained 30 image sets, and 1000 image sets. The only difference was how long it took to get the LORA trained.

It depends on what you are after, in my experience likeness/character don't require as many images, but requires GOOD images. Style/concept takes a lot more images but they don't need to be as high quality.

You can select a batch size, see if you go OOM, adjust up or down from there. I like to leave a little room for overhead.

Good luck!