Datasets › XYSquares

XYSquares

Introduced by Nathan Michlo et al. in Overlooked Implications of the Reconstruction Loss for VAE Disentanglement1 Jan 2022 archive 2025-07-28

Synthetic dataset intended for benchmarking disentanglement frameworks.

XYSquares is adversarial in nature; the distance between any two observations in the dataset is constant when measured using a pixel-wise distance function. It is usually impossible for VAE frameworks that use pixel-wise losses to disentangle this dataset.

The dataset is constructed from 3 non-overlapping red, green and blue squares that are each 8×8 pixels in size. Each square can move along the x and y coordinates of an 8×8 grid. The resulting images are 64×64 pixels in size. With this construction the dataset has a total of 8 ground-truth factors for a total of 8⁶ = 262144 observations.

Benchmarks archive 2025-07-28

No leaderboard in the archive resolves to this dataset.

Papers archive 2025-07-28

No paper in the archive has a leaderboard row on this dataset; the archive counts 1 paper for it but never published that list.

Dataset loaders archive 2025-07-28

1 loader as listed in the archive; links are outbound and not re-checked here.

Tasks archive 2025-07-28

License archive 2025-07-28

MIT

Modalities archive 2025-07-28

Languages archive 2025-07-28

No language tagged.

Variants archive 2025-07-28

  • XYSquares

1 variant name, as the archive lists them.

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections