10 upvotes
Wiener Estimates Speed Up Pixel-Flow Model Training
What is Beyond the Estimate: Wiener-Informed Decoding for Easier Training of Pixel-Space Flow Models?
The paper shows that inserting frozen Wiener estimates, simple linear predictions derived from coarse image statistics, helps neural networks learn pixel-space velocity fields faster and more reliably. The improvement extends beyond the linear predictions, boosting final flow quality without extra training cost.
Who it is for
Anyone training optical-flow or video synthesis models can reduce training time and improve results by using this Wiener-informed decoding trick.