How to train my own lora checkpoint?

[HZIH]

I’ve tried to train a stable diffusion xl lora with diffusers official example, however, the result is bad and the size of the lora checkpoint is much smaller than the one provided such as Harry potter 1.8G.

Any advice?

[HZIH]

I've tried to train sdxl lora with 15 images with sd-scripts, however, the result is so bad, any advice?

[kongzhecn]

I followed the tutorial in this video: https://www.youtube.com/watch?v=y2J7EZUk_a0 for reference. You might find it helpful as well.

[HZIH]

Thx, I've tried the method in this video but the result is bad. And when I integrate the trained checkpoint into OMG, the generated image is not satisfactory and even worse.

Moreover, civitai.com cannot provide checkpoint of celebrity now. I used to try the checkpoint on civitai, it's much better than the model trained in this video https://www.youtube.com/watch?v=y2J7EZUk_a0

[HZIH]

Here are some examples: