Show HN: ReflexConv2d – 57% less blur in image reconstruction

Category: library

Tags: convolution, image-reconstruction, pytorch

Score: 5.5/10 (Innovation: 6, Technical: 6, Documentation: 6, Utility: 4)

ReflexConv2d is a PyTorch convolutional layer that uses a self-modulating spatial mask derived from its own weights to reduce blur in image reconstruction tasks like U-Nets. It offers a novel approach to spatial filtering with clear benchmark improvements, though its practical utility is currently niche due to limited adoption and testing.

Target audience: backend devs, data engineers

Repository: https://github.com/singam96/ReflexConv2D · Python · Apache-2.0

View on Hacker News