Size of the dataset?

#1
by Viewegger - opened

Thank you for making this public, could I ask you about the size of the dataset you used for finetuning?

15 images

HL2Tex_00008_.png
HL2Tex_00010_.png
HL2Tex_00011_.png
HL2Tex_00012_.png
HL2Tex_00013_.png
HL2Tex_00015_.png
HL2Tex_00019_.png
HL2Tex_00022_.png
HL2Tex_00023_.png
HL2Tex_00024_.png
HL2Tex_00025_.png
HL2Tex_00031_.png
HL2Tex_00033_.png
HL2Tex_00036_.png
HL2Tex_00037_.png

thank you for such a quick answer and even more for providing examples, I was thinking whether I can train model to generate seamless pattern that could be repeated and when I have seen you LORA I thought that it is supposed to achieve that.

But then looking more closely the results aren't seamless. But seeing the dataset it's no surprise :) Because many of the images are not seamless.

Thank you nonetheless!

Viewegger changed discussion status to closed

image.png

image.png

image.png

This is tiling test, i think what would be cool is you can try to cherry pick good ones and train another lora with good examples from this LoRA.

Maybe you didnt zoom to examples that is why you think they are not seamless. Huggingface shows them square but my examples are 1024x768 thats why you might think they are not seamless.

This is tiling test, i think what would be cool is you can try to cherry pick good ones and train another lora with good examples from this LoRA.

Yes but I was thinking about more complex pattern at least like the snowflakes you have in results, I will try during next week probably...

Yes, i also recommend using good captions. It is really easy to make it for 15 images with hand, you dont need autocaptioner vlms. They would be more accurate and rich.

"smlstxtr, A close-up view of lava with various shades of orange and black, seamless texture"

"smlstxtr, A blue surface with square indentations and screws, seamless texture"

My captions were like this they are not that much descriptive

Sign up or log in to comment