Thanks for sharing this result and the candid feedback. We certainly have a long way before the model can approach human quality. The fluting issue from the p

Thanks for sharing this result and the candid feedback. We certainly have a long way before the model can approach human quality. The fluting issue from the pillars, for instance, presents a dilemma since the model is trained to consider these subtle textures as noise. Besides improving the Microsoft model, we're also researching our own model that hopefully will yield better results.