Wiener-Informed Decoding Improves Pixel-Space Flow Model Training

Wiener-Informed Decoding (WiD) uses frozen second-order image statistics and a trainable Wiener prediction path to stabilize velocity learning in pixel-space flow models. The method injects statistical knowledge into existing decoders via coefficient interfaces, improving training efficiency without requiring modifications to the native model blocks.