CoolFace
Apppublic

mashhoodnnjjjj/masked_autoencoder

sourceHugging Facemitupdated 7mo agoView on Hugging Face
0likes
App README

🎭 Masked Autoencoder — Image Reconstruction

Upload any image and watch a Vision Transformer-based MAE reconstruct it from only a fraction of the visible patches.

Model Architecture

ComponentDetails
Encoder12-layer ViT · 768-dim · 12 heads
Decoder12-layer Transformer · 384-dim · 6 heads
Patch size16 × 16 px → 196 patches per image
Input size224 × 224
Trained onTiny ImageNet 200

How to Use

  1. 1.Upload a JPG or PNG image
  2. 2.Adjust the Mask Ratio slider (default 75%)
  3. 3.Click Reconstruct
  4. 4.Compare Original → Masked Input → Reconstruction side by side