Why does my kohya LoRA work well in SD1.5 but not in other models?

I'm new to training so maybe I'm missing something obvious.

I'm trying to train an art style (coloring pages).

Here are my data set:
https://drive.google.com/drive/folders/1QtDTMMG2uGc10xI9Vr8E5Kh9x7Dun9FZ?usp=sharing

I've trained in kohya using all the default settings except for these settings (just some random recommendations from guides)

- class prompt: coloring page
- instance prompt: sks style coloring page
- Network rank 128
- Disable transformers
- Epoch 10, repeat 40, save every 1 epoch
- I didn't use regularization images (I dont care if it overfits I will only use it to generate this kind of images).
- I didn't use captions (not sure if I should?)

Here's the output for all the steps for the 1.5SD model
https://drive.google.com/file/d/13BJgIg0CjlUrQ1MiVD_Li0i_ka6N_0PY/view?usp=share_link

Here's the output for all the steps for the Deliberate v2 model
https://drive.google.com/file/d/1GPb0ITK9g5dZ4J5w8XEvBbqLEMdOStKW/view?usp=share_link

As you can see, in the deliberate base model, the effect is very small. In fact almost none for the tiger and sports car. Some effect for the turtle and planet.

Am I using Kohya incorrectly?

The issue names no source file, test, or entry point. Start by reviewing the linked dataset, training outputs, and listed kohya settings; done would require determining whether the model-dependent behavior is an sd-scripts defect or expected training behavior and documenting the conclusion.

Written by the indexing model from the issue text.