Intern, Warmup for Hybrid Implicit Neural Representation Compression

Summary Hybrid implicit neural representations (INR) compress images by combining learnable feature grids with lightweight neural networks.

The grids capture spatial information at multiple resolutions, while a compact network transforms these features into pixel values.

Both components are jointly optimized for each image to balance bitrate and reconstruction quality.

The resulting latent features and network parameters are quantized and encoded into a compressed bitstream.

During decoding, th…