Will be grateful for you help about LORAs
Your LORAs are really good! For several days I'm trying to make a silhouette Lora for my project and failed miserably after testing many approaches.
I've made a 170 images dataset, several types of descriptions etc, made a quite a few training's but results not good.
I would be so appreciated if you can share your Lora training config (I'm using Kohya SS and SDXL 1.0) and several images/prompts pairs so I understand how it build. Any other tips and tricks will be much appreciated.
Thank you very much in advance!
No hidden secrets. Nothing secret. Caption normal like all people do. Normal settings from a lot of place work well too.
After so many days it seems like, I don't know what is "normal" here as it somehow doesn't work for me :)
- Could you please show couple of image descriptions of your coloring book Lora? I'm sure I just doing them wrong
- I see you've used 32 Network Alpha and Dim. What learning rate works best for you? (if you can share your exact settings it would be much easier :) )
I dont have my images and prompt for that rn.
I dont have the settings used for that specific Lora.
Normal settings and prompts are what you can find among the hundreds of tutorials on the internet. There is no secret. This is a good start. The rest is small tests and changes. Datasets, configurations. SD is like that. Trial and error. There's no secret. There is no universal setting, although the normal settings will deliver a good result for 99% of people. You just need to test if you need a much better lora. Then you go for a trial and error.
I am trying to save the model locally. I am having some issues in trying to save it from the .safetensors file as there in no model_index.json file for information about the model. Can you show me how I can use your LoRA using transformers directly?