So I'm trying to do DreamBooth on an SDXL anime model, and I'm wondering—do regularization images actually help or not?

The Forge / 工坊 Models / LoRA / 模型与 LoRA
1

Someone asked me to test whether regularization images (aka class images) actually make a difference in SDXL DreamBooth training, so I used an anime SDXL model as a baseline. The author said they ran over 120 full training runs before coming to a conclusion, using Kohya SS GUI.

Their setup: for the group with reg images, 1 epoch, repeat 150, save every 30 epochs; for the group without reg images, repeat set to 1, train for 150 epochs, save every 30 as well, so the model naming format is different—one marks epochs, the other marks steps.

They also admitted the training set is only mid-tier quality, with repeated backgrounds and outfits, and missing full-body poses, which kinda messes with how reliable the conclusion is. From a workflow design perspective, I’d rather see if the control variables are clean and whether the step counts between the two groups are actually equivalent.

Anyone who’s run a comparison—is it worth the hassle to use reg images on anime models?

2

The training set quality is trash, so take the results with a grain of salt.

3

Regular maps work for realism, but for anime it’s hit or miss.

4

The key is whether the two sets of steps are equivalent or not.

5

Running it 120 times is some real dedication.

6

Kohya’s got too many parameters, easy to mess up.

7

Ugh, the naming format is all over the place, such a headache trying to match them up.

8

Damn, that’s impressive.

9

agree, regularization images only show their difference when the training set is clean, mix in dirty data and everything falls apart