stegastamp-original

StegaStamp Original

A PyTorch implementation of StegaStamp (Tancik et al., CVPR 2020) trained on COCO dataset.

Model Details

• Architecture: StegaStamp encoder-decoder with Spatial Transformer Network (STN)
• Training: 140,000 steps on COCO train2017 images
• Input size: 400×400 RGB images
• Message capacity: 100 bits

Performance

Training Configuration

experiment_name: stegastamp_original
data:
  batch_size: 4
training:
  num_steps: 140000
  lr: 0.0001
  stn_lr_scale: 0.01  # Critical for training stability
loss:
  message: { scale: 7.0, ramp_steps: 1 }
  l2: { scale: 2.0, ramp_steps: 10000 }
  lpips: { scale: 1.5, ramp_steps: 10000 }
distortion:
  strategy: curriculum

Key Training Insight

The STN linear parameters (stn_fc_weight, stn_fc_bias) require a 100× lower learning rate to prevent training collapse. Without this, the decoder outputs constant 0.5 probabilities (loss stuck at 0.693).

Usage

import torch
from picode.models.stegastamp import Encoder, Decoder

# Load checkpoint
checkpoint = torch.load("best.pt", map_location="cpu")
encoder = Encoder(num_bits=100)
decoder = Decoder(num_bits=100)
encoder.load_state_dict(checkpoint["encoder"])
decoder.load_state_dict(checkpoint["decoder"])

# Encode message into image
image = ...  # (B, 3, 400, 400) tensor in [0, 1]
message = torch.randint(0, 2, (B, 100)).float()
encoded_image = encoder(image, message)

# Decode message from image
logits = decoder(encoded_image)
decoded_message = (torch.sigmoid(logits) > 0.5).float()

Citation

@inproceedings{tancik2020stegastamp,
  title={StegaStamp: Invisible Hyperlinks in Physical Photographs},
  author={Tancik, Matthew and Mildenhall, Ben and Ng, Ren},
  booktitle={CVPR},
  year={2020}
}

License

Apache 2.0