Using my own Comics as Input

#467
by GiacomoGeicher - opened

I duplicated the space and connected it to a CPU. But it does not work with my data input (jpg.). What kind of format should my sketches be in? Even with the suggested test-set (diffusers/dog-example) it does the same (load, train, error).

I would really appreciate any kind of help! Thank you in advnace!

Bildschirmfoto 2023-12-05 um 16.01.45.png
Bildschirmfoto 2023-12-05 um 16.02.15.png

Hi,

  • I know sometimes it displays an "error" or a "timeout", but it is actually working: if you go look into the logs, you see it is processing etc (this may happen if the number of training steps is long, and the training lasts for hours)
  • may be an issue with the dataset which perhaps wasn't completely uploaded before you went to the next step (there is a bit of a wait time)
  • normally I think it works with .png, .jpg, .jpeg.. but perhaps there is an issue with the file name or file path (did you drag & drop a folder or directly the images?)

I've posted your issue in the community forum of the creator of the SDXL LoRA trainer:

https://huggingface.co/spaces/fffiloni/train-dreambooth-lora-sdxl/discussions/4

P.S. once it works, you may have to experiment with the number of steps and images to get good results (personally I tried with 10 images, 50 images.. and training of 500, 3000.. I don't necessary recommend large values as this costs money and quality improvements are not guaranteed - there can be side effects, like overfitting to specific poses or image layout etc - but yeah often a bit of trial and error is needed to improve quality, at least at first)

Sign up or log in to comment