Skip to content

DySample

A lightweight dynamic upsampler built on learned sampling

Visit official site

What DySample is

The method is described as a lightweight and efficient dynamic upsampler that learns upsampling from a sampling perspective, avoiding time-consuming dynamic convolution and extra sub-networks.

What you can do with it

  • Learn upsampling from a sampling perspective
  • Skip dynamic convolution and extra sub-networks
  • Run a fast implementation without a custom CUDA package
  • Keep the parameter count and FLOPs low

Application scenarios

  • Serve as an upsampling module in semantic segmentation models
  • Act as a feature extraction module in object detection models
  • Serve as an upsampling module in depth estimation models

Target fields

  • Image processing work
  • Computer vision tasks

Pros & cons

✓ What we like

  • A fast implementation that skips extra sub-networks
  • Dynamic convolution is avoided in the design

! What to watch out for

  • The design replaces dynamic convolution with learned sampling
  • The material names five tasks for the comparison with other upsamplers

FAQ

Does it need a custom CUDA package?

The feature list states the fast implementation does not rely on an extra CUDA package.

Which tasks are named?

Semantic segmentation, object detection, instance segmentation, panoptic segmentation and monocular depth estimation are listed.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review