Fashion Autoencoder — three views
Three trained autoencoders on Fashion MNIST 28×28 grayscale (T-shirts, sneakers, dresses, etc.). Pick a tab below to explore reconstruction, denoising, or latent-space interpolation.
Scope: the models were trained on the 10 Fashion MNIST classes at 28×28 grayscale. They're optimized for that distribution — they will not produce useful reconstructions of arbitrary photos. Pick one of the bundled samples to see what the models do well.
Pick a sample (or upload). The two trained autoencoders compress the 784-pixel input down to a small latent and reconstruct it. Compare how each architecture loses or preserves detail.
Add Gaussian noise to a sample and watch the denoising autoencoder recover the clean version.
| Input | Noise factor (σ) |
|---|
Encode two samples to the convolutional latent space, then decode points along a straight line between them. The middle frames are not real garments — they are what the model imagines in between.
| Sample A | Sample B | Interpolation steps |
|---|