about own datasets

[fumoab]

Hello, thank you very much for your work on the stylized dataset. As a beginner in this field, I want to conduct experiments by training with my own private dataset. I plan to create my own content images and manually create their content text descriptions. For the style aspect, I will use your StyleGallery.jsonl file. Could you please provide me with some steps or suggestions?

[Jeoyal]

Hi @fumoab , thank you for your interest in our work. What are your content images look like?

[fumoab]

Hi @fumoab , thank you for your interest in our work. What are your content images look like?

Thank you for your reply. My content images are water meters in various styles, and I have resized them to 320*320. Additionally, I manually wrote text descriptions for them and created corresponding txt files, with a directory structure like:

image
--000001.jpg
--000002.jpg
...
txt
--000001.txt
--000002.txt
...

I suspect I might need to adjust them further (resize to 512*512) and add them to StyleGallery.jsonl (setting them as styleless?) or adjust them and train on the basis of StableDiffusion v1.5.

[Jeoyal]

Could you please post some of your content images on this issue window?

[fumoab]

Could you please post some of your content images on this issue window?

no problem!

[Jeoyal]

What will your final stylized image look like?

[fumoab]

What will your final stylized image look like?

These are images I generated using traditional style transfer techniques.

[Jeoyal]

I suggest you to inference on our StyleShot_lineart, a version taking the lineart of content image as highly detailed control.

[fumoab]

I suggest you to inference on our StyleShot_lineart, a version taking the lineart of content image as highly detailed control.

Thank you very much, I'm going to try it!