Region-Adaptive Generative Compression with Spatially Varying Diffusion Models

A research team from ETH Zurich and Disney Research has developed a breakthrough AI-powered image compression system. Their paper, 'Region-Adaptive Generative Compression with Spatially Varying Diffusion Models,' introduces a codec that fundamentally changes how bits are allocated within an image. Instead of treating all pixels equally, their novel diffusion model applies varying amounts of denoising per pixel based on an 'importance map.' This allows the system to dedicate more representational capacity to salient regions—like a person's face in a photo—while spending fewer bits on less critical backgrounds.

The core innovation is a 'spatially varying diffusion model' capable of following these arbitrary importance maps, which act as a prior for the compression process. The team further integrated these maps into the entropy model, improving the crucial rate-distortion performance. The result is a codec that outperforms current state-of-the-art region-of-interest (ROI) controllable baselines in both full-image and ROI-masked perceptual quality tests. This means reconstructed images look more realistic and detailed, especially in the areas humans naturally focus on, without proportionally increasing file size.

This approach directly targets a key inefficiency in standard and even generative compression: human vision is highly non-uniform. We don't perceive every part of a scene with equal detail. By building a model that understands and exploits this property, the researchers have created a path toward much more efficient visual data storage and transmission, promising clearer images and videos at significantly lower bandwidth costs.