Hung Tran, a master's student at the University of Information Technology (VNU-HCM), has published code and data for his paper "The Clipping Trap: A Silent Failure Mode in Overfitted Neural Image Codecs." The work highlights how overfitted codecs like Cool-chic and C3 can produce flat color images with zero distortion gradients on specific content types.

  • Overfitted codecs clamp the decoder output inside the differentiated loop, causing the distortion gradient to become exactly zero on bright, low-contrast photos and screen content.
  • The encoder reduces rate without raising an error, resulting in outputs that PSNR, SSIM, MS-SSIM, and LPIPS score no higher than a constant image.
  • A straight-through estimator removes this failure with only ≤0.1% BD-rate impact on the Kodak dataset while keeping the bitstream identical for unmodified decoders.

The author is seeking an arXiv cs.CV endorsement (code KGU7X3) and welcomes feedback or criticism on the paper.