Back to projects 
Angel Diffusion
[ December 2022 ]
I finetuned Stable Diffusion on my dog Angel. 
Below are some generated images of Angel. I think the results are meh. The variance of output quality is pretty wide. Some have a close resemblance to Angel and others not so much. I underestimated how much hyperparameter tuning you had to do to go beyond just average results and produce a really excellent finetune. So I began fiddling with certain hyperparameters such as max_train_steps, num_inference_steps, and guidance_scale. The output quality turned out to only be marginally better, though. Perhaps a better tuned prompt would fix the images that turned out to be duds. I'm going to continue tweaking the hyperparameters and see where it gets me...
Because I don't have the GPU resources locally to run the finetuning and model inference process, I ran my code on a rented A100 in the cloud using a new serverless startup called Modal. It cost about a dollar to finetune on the rented A100, and a few cents to run inference, but it's all well within Modal's generous free tier. I really like the developer experience of using Modal and the effortlessness of making my code run on their GPUs, so I'm going to keep using them to run my serverless compute jobs.