Bad Results with Example Config

[BieHDC]

The default example here gives me very bad and mangled results. Replacing the linked sdxl model to this one fixes the issue for me.
Does it need the commented out vae enabled or is the base model this links to bad?

I noticed this when i tried to learn to make a lora and eventually plugged in a different sdxl model because i tried everything else to get a good result and it worked immediately.

Can i also make a more complex and involved example and donate it to this repo? The current examples are very bare bones and i would like to share a more real project to help others learn. And also something anyone can take and just run for training to get a good known working example to work off of. I, of course, will take the images myself of a real world object and base the lora on that to not have any licensing issues.

[hipsterusername]

Interesting- these should be the same! Thanks for reporting it.

We would definitely appreciate a contribution to help folks learn and have better examples :)

[BieHDC]

We would definitely appreciate a contribution to help folks learn and have better examples :)

I finally was able to come back to this after - checks calender - over a month and finishing my custom dataset manager for a better experience.
However in the meantime you made a video about the topic. It seems you actually would have everything required on hand already.
Would you share the config.yaml files the training program generates? I understand that the interface/tool itself is for the enterprise version, but i assume the backend is this training program and we could just use those as a base. This would also save me a lot of time of testing parameters (and gpu thermal paste).
I still have yet to pick out a real world photographed object for an example, but i already got a video game character and style picked out as example. Specifically Saints Row 3 because the character editor allows for different outfits and close up zooms for details which yielded good initial results.