[Rafael Bravin]

Hello, I’ve just created a Lora model using the tutorial . At first I got the following error:

modules.devices.NansException: A tensor with all NaNs was produced in Unet. This could be either because there’s not enough precision to represent the picture, or because your video card does not support half type. Try using –no-half commandline argument to fix this.

 

Then after trying the suggested procedure with no success I added –disable-nan-check to the arguments and now it doesnt throw any errors but the images turn all black

 

[Andrew]

Hi, would you be able to follow the training example and reproduce the result? The first step is to make sure your training process is correct.

The notebook is recently updated — Did you use the newly updated one, or the koyha_ss (gradio-based) one? Did you train SDXL or SD 1.5?

[Rafael Bravin]

I used the notebook referenced in the resources page, and followed all the steps described in the tutorial, the difference was that I made my own images to use as a dataset, and that I was using 100 images. I was also doing the steps and following with my own computer and it worked, but I  made a few adjustments since the first one went wrong, locally I used 25 images and chose Lycoris/LoKr as the Lora type. I suspect it was something about the number of images or something about the notebook, or even a wierd error during training.

I will create another topic to talk about another subject, please give the other topic preference to help me

[Author]

Posts