Beyond the Estimate: Wiener-Informed Decoding for Easier Training of Pixel-Space Flow Models

Benchmark Model Rank Results
image-generation-on-imagenet-512x512PixelDiT-XL+WiD#16FID: 1.64
image-generation-on-imagenet-256x256PixelDiT-XL+WiD#39FID: 1.53
text-to-image-generation-on-dpg-benchDeCo-XXL + WiD–DPG-Bench Score: 82.6
text-to-image-generation-on-genevalDeCo-XXL + WiD–Overall: 0.86