As an Amazon Associate we earn from qualifying purchases.
I think the current StyleSynthesis benchmarks don't accurately reflect real-world performance.They frequently enough focus on very specific,controlled conditions with carefully curated datasets.This makes the systems look amazing in the publications, but when you try to apply them to more diverse and challenging content, the results can be disappointing. It feels like we're optimizing for the benchmarks themselves, rather than building truly robust and generalizable StyleSynthesis tools.
For exmaple, many benchmarks use datasets with perfectly aligned subject and style examples. In real life, you rarely have that luxury. You might want to synthesize a photo with a painting style, but the painting wasn't created specifically to match the photo's composition. This difference in the data drastically affects the quality of the output.
Are others experiencing similar issues? It feels like we need benchmarks that better capture the variations and complexities of real-world style synthesis tasks. What alternative evaluation metrics or datasets do you think would be more informative?