Many settings affect the outputs in interesting ways, but that's half the fun. These LoRAs are very lightly trained; more training may or may not help. The competitions are also performed using zero-shot text guessing, and if Facebook said it, you can bet that's actually Meta AI saying it, and they are leaders in the field. Source: about 1 year ago

You should look at the entire niche of MAE-related papers, that's quite exciting, and the neuroscience-inspired stream of stuff like Barlow Twins. As well, the official Facebook AI blog is surprisingly good coverage of much of the interesting un/semi-supervised DL research FAIR does, and worth going through. Source: almost 2 years ago

At Scale we built a tool for model debugging in computer vision called Nucleus (scale.com/nucleus) designed exactly for this, which is free try out if you're curious to see where your model predictions are most at odds with your ground truth. Source: over 2 years ago

To address your point about gathering edge cases, which can also be defined as cases of low model fidelity for our use cases, there is active learning and tools such as Aquarium Learning and Scale Nucleus which make it easy to implement into workflows. Source: almost 3 years ago