CC-4DGS reduces dynamic Gaussian Splatting storage to 20–30 MB per scene

A paper titled CC-4DGS has been published on arXiv, proposing a way to reduce the storage footprint of dynamic 4D Gaussian Splatting scenes from tens to hundreds of megabytes down to 20–30 MB. The authors replace large trainable hash tables with a computational deformation field using deterministic dense hash encoding and compact neural network decoders, and also compress canonical point cloud attributes through conditional autoencoding, selective quantization, and residual codebooks. As described by the author, reconstruction quality is comparable to Swift4D with significantly better storage efficiency. The code has been released as open source. More details in the paper on arXiv.

What it means

This news relates to Gaussian Splatting and compression of dynamic scenes, not to 3D generators from images. In our section on AI for 3D, we currently feature tools of a different class — image-to-3d generators — and CC-4DGS does not directly overlap with them. The closest tool pages in topic are OpenLRM, Hunyuan3D 2.1, and TRELLIS — they solve the task of generating static 3D models from a single image. None of them work with 4D Gaussian Splatting or use the deformation or point cloud attribute compression mechanisms described in CC-4DGS. The fact sheets for these tools have no fields about scene storage or Gaussian attribute compression, so a storage size comparison cannot be made. For those working with dynamic scenes and Gaussian Splatting, CC-4DGS may be of interest as a way to reduce storage usage without losing real-time rendering. For users of 3D generators from our reference, this work does not yet change either the requirements or the capabilities of existing tools.