Squeeze3D: Extreme Neural Compression with Latent Space Bridging
Squeeze3D introduces a framework to achieve extreme compression of 3D data by bridging latent spaces between pre-trained encoders and generative models. It compresses input data into a compact latent code using trainable mapping networks, which are trained on synthetic paired data. This allows existing models to be reused for compression across various 3D formats without needing specialized training for each object.