Doesn't always work, but gets you a lot closer quickly compared to just trying keywords.
Dataset of 18 images, 3 vectors, batch size of 3, 3 steps gradient accumulation, trained for 500 steps or 250 epochs.
-