MegaStyle

[404-brittle]

The recent development of ControlNet-LLLite v2 has opened the possibility of using style-transfer datasets such as:
https://huggingface.co/datasets/tencent/MegaStyle-1.4M
After successfully training Canny/Lineart:
https://civitai.com/models/2604674/controlnet-lllite-anima-preview-3
It was my understanding that this technology could succeed for style transfer. However, it doesn't seem to work well.
Is this a fundamental limitation of ControlNet-LLLite or perhaps there needs to be changes to hyperparameters?
I've tried a number of different settings and so far the resulting style transfer LoRA does not learn much at all.
Any input would be well received.

[kohya-ss]

Unfortunately, I believe LLLite is only effective for controlling by the positional information of the control image and the target image are mapped. In other words, it's not suitable for extracting abstract information like style.

This is because LLLite is very lightweight, and each coordinate in the control image is directly mapped to each coordinate in the target image.

Style transfer would require a network with a different architecture.

[404-brittle]

I appreciate the confirmation.
If there is future work for adding more complex networks I will happily run the experiment again.
Thank you