Why does the model of first stage for finetune train all parameters instead of using LoRA?

fcjian / InstaGen

InstaGen: Enhancing Object Detection by Training on Synthetic Dataset, CVPR2024

https://fcjian.github.io/InstaGen

MIT License

68 stars 2 forks source link

Open ohheysherry66 opened 4 months ago

ohheysherry66 commented 4 months ago

Thank you for your great job! The question is described in title.

fcjian commented 2 months ago

@ohheysherry66 You can also fine-tune SD using LoRA. We tried it at the time and it achieves similar performance.