Wiener-informed decoding speeds pixel-space flow model training
Original title: Beyond the Estimate: Wiener-Informed Decoding for Easier Training of Pixel-Space Flow Models
On ImageNet 256, PixelDiT-XL with WiD reports FID 1.53 after 500 epochs, matching prior quality with 37.5% fewer epochs.
The paper argues that frozen Wiener estimates and related statistical structure make learning pixel-space velocity fields easier. Wiener-informed Decoding adds a trainable Wiener path and coefficient interfaces that feed a native decoder without changing its blocks, with parameters trained under the original velocity objective. The authors report faster training, better generation, smoother gradients, and modest extra parameters across scales and decoder designs, including class-conditional and text-to-image settings.
- PixelDiT-XL with WiD reaches FID 1.53 on ImageNet 256 after 500 epochs.
- Gains are also reported from frozen bases and pixel decoding.
- The method stays effective when used with representation alignment.
- All new parameters, including spectrum and coordinate transforms, are trained jointly.
From the original article
1 min read at huggingface.co
Learning velocity fields directly in pixel space is difficult and slow. We show that coarse second-order image statistics can ease this task: frozen Wiener estimates improve velocity learning, with gains extending beyond the linear predictions they explicitly provide. Moreover, we observe similar but complementary gains from frozen bases and pixel decoding.
By Ziqiao Wang, Yang You · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.