Skip to content

DeepLens

DeepLens is a PyTorch-based differentiable optical lens simulator for end-to-end computational imaging, supporting multiple optical models — geometric ray tracing, diffractive wave propagation, hybrid ray-wave, and surrogate PSF networks.

DeepLens supports three main use cases:

  • Differentiable optics — accurate and efficient gradient computation for optical parameters via differentiable simulation and backpropagation.
  • Automated optical design — fully automated optical design driven by gradient information and advanced optimization algorithms.
  • Computational imaging — physically accurate image simulation and end-to-end optimization with downstream image reconstruction algorithms.

DeepLens computes the point spread function (PSF) of an optical lens in a fully differentiable manner:

Light source → [ DeepLens ] → PSF

DeepLens also serves as the differentiable optics engine in an end-to-end computational imaging pipeline such as End2endImaging, where the optics, sensor, and a reconstruction network form a single differentiable graph that can be optimized jointly:

Scene Image → [ DeepLens ] → Spectral Image → [ Sensor ] → Raw Image → [ Network ] → Output Image
                │                               │                        │
                GeoLens                         RGBSensor                UNet
                HybridLens                      MonoSensor               Restormer
                DiffractiveLens                                          NAFNet
                ParaxialLens
                PSFNetLens

Key Features

  • Differentiable Optics — differentiable optical simulation for accurate, efficient gradient computation, enabling lens inverse design.
  • Automated Design — fully automated optical design via gradient-based and advanced optimization, shortening the development cycle for a wide range of systems (highly aspherical lenses, metasurfaces, AR/VR displays).
  • Multiple Optical Models — geometric ray tracing (GeoLens), hybrid ray-wave (HybridLens), pure diffractive (DiffractiveLens), neural surrogate (PSFNetLens), and defocus (DefocusLens) models.
  • Image Simulation — photorealistic rendering with spatially-varying, depth-dependent aberrations, closing the sim-to-real gap when combined with End2endImaging.

Advanced Features

Additional capabilities, customizable upon request:

  • GPU Kernel Acceleration — custom GPU kernels deliver >10× speedup and >90% memory reduction across NVIDIA and AMD platforms, making deployment on local laptops practical.
  • Polarization Ray Tracing — polarization ray tracing and inverse design of thin films via DiffTMM.
  • Non-Sequential Ray Tracing — differentiable non-sequential ray tracing for stray-light analysis and optimization.
  • Distributed Optimization — distributed simulation and optimization for billion-scale ray tracing and high-resolution (>100k × 100k) diffractive propagation.

Getting Started

  • Setup

    Create the environment and run your first lens demo.

    Set up DeepLens

  • Quickstart

    Load a lens, compute a PSF, and render an image in minutes.

    Get started

  • API Reference

    Full class and function documentation for the optics engine.

    Browse the API

  • Examples

    Lens design, end-to-end optimization, and image simulation.

    See examples