Hi, thank you for nice work! And I want to know do you try finetune stable diffusion (v1.5 / v2)on mnist instead of training directly on mnist, I have tried, but the result is bad, the pre-trained model maybe can't recognize digits. May I kindly ask for your suggestions?