Fine-tuning guide
Hi, I am fascinated by your model. Can you please describe the fine-tuning steps? Such as number of instance images, class images, learning rate, number of steps, the prompt, and so on... There is no much finetuning done one the specific style.
Hello,
For Analog Diffusion I used around 110 images. My quick count has it at 30% closeup-medium shots, 25% medium-full shots, and the other 45% as non-people or images where no face is visible (such as silhouetted or facing away). It was trained with prior preservation, 1500 class images made with the prompt "photograph", LR of 1e-6 with polynomial scheduler to 10k steps (batch size 1), and trained at 512x512. Since it was class training with ShivamShrirao's repo, no captions needed.
Hello, I am fascinated by your work and was curious about the dataset you used. Did it have pictures of the same person or did it have pictures of many random people?
What was the source of the 110 images? Mind sharing? I want to train a similar model for SD 2.1
First off amazing work! Your style models are fantastic.
You mentioned you used ShivamShrirao's repo. Is it the one found here? https://github.com/ShivamShrirao/diffusers/tree/main/examples/dreambooth