Training Lora based on PixArt-XL-2-1024-MS

[Sainzerjj]

I am training Lora on PixArt-XL-2-1024-MS using the code from this repository, using text prompts like "1girl, solo, short hair, black hair, hair ornament, dress, bare shoulders, brown eyes, jewelry," and found that although I set the learning rate very large, it still need to train 1000+ steps, I want to ask if this is normal or is there an additional training technique?

[lawrence-cj]

This may be caused by the small graident clip.