Nothing beyond the standard advice for SQLite performance. I found that keeping the data in individual per-image database files worked a lot better than trying to create one big database for the whole stack. With that done, I expect the big performance win to come with parallelization of the image transformation tasks.