← All Cracked Math projects
PM-A07NOT STARTEDADVANCEDMATHPYTHONIMPACT ★★★☆☆

2D Gaussian splatting with every gradient derived by hand

No autodiff: apply the chain rule by hand through 30,000 anisotropic Gaussians and fit a Kodak photo.

EST.30 h
REWARD+400 XP
AFTERZhang et al. 2024 + 1

ML-ARENA BENCHMARKS

PROOF: BENCH
METRICINDUSTRYMINE
Mean PSNR (peak signal-to-noise ratio, an image-quality score in dB; higher is better) over the 24 Kodak images at 30,000 Gaussians; finite-difference gradient check
GOAL · Hand gradients match finite differences to 1e-5 relative; mean Kodak PSNR within 2 dB of GaussianImage's reported number at 30k Gaussians
GaussianImage (arXiv 2403.08551) Kodak results at the same Gaussian count (GPU) · compared against their reported tables—
no results yet
YOUR RESULTS

Your recorded results appear here.

Brief

Represent each Kodak image (from a standard set of 24 test photos) with N anisotropic 2D Gaussians, meaning stretched and rotated blobs. Each Gaussian has a covariance parametrised by its Cholesky factor (so it always stays valid), a colour and an opacity. Write a tiled, vectorised NumPy rasteriser (it draws the Gaussians into pixels, one screen tile at a time) and its backward pass, derived by hand through the 2x2 inverse, the exponent and the accumulation; no autodiff in the solution path. Train with your own Adam plus densify/prune heuristics (add Gaussians where the error is high, remove those that contribute little).

VICTORY CONDITIONS

  1. Hand gradients match finite differences to 1e-5 relative; mean Kodak PSNR within 2 dB of GaussianImage's reported number at 30k Gaussians (Mean PSNR (peak signal-to-noise ratio, an image-quality score in dB; higher is better) over the 24 Kodak images at 30,000 Gaussians; finite-difference gradient check).
  2. A public repository with a README that shows your result and how to reproduce it.

Terms in plain words

Gaussian splatting
Representing an image or scene as many soft Gaussian blobs that are drawn and blended together.
Anisotropic Gaussian
A Gaussian blob that can be stretched more in one direction than another and rotated, set by its 2x2 covariance.
Cholesky parametrisation
Storing a covariance as L L^T with L lower triangular, so any values of L give a valid covariance.
Rasteriser
The code that turns shapes into pixel colours.
PSNR
Peak signal-to-noise ratio, a log-scale measure in dB of how close a reconstructed image is to the original.
Densify / prune
Heuristics that add Gaussians where the image is poorly fit and remove ones that barely contribute.

Why it rates ★★★☆☆

Hand-derived Jacobians through a real renderer prove matrix-calculus fluency, though 2DGS (2D Gaussian splatting) itself is published.