Bridging the Gap between Gaussian Diffusion Models and Universal Quantization for Image Compression

Figure 1. Visualization of the 3 gaps we address in this work. Failure to match the noise level (middle columns) results in either too noisy or too smooth images. Inconsistent noise types (middle-right) introduces generative artifacts and color shift. Applying diffusion to discrete data (far right) causes flat textures as well as color shift. Addressing all three gaps (middle-left) results in the most realistic reconstruction that best matches the source image (far left).