DiffCom: Decoupled Sparse Priors Guided Diffusion Compression for Point Clouds

Xiaoge Zhang, Zijie Wu, Mingtao Feng, Mehwish Nasim, Saeed Anwar, Ajmal Mian

The University of Western Australia, Xidian University

IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2026

DiffCom is a decoupled sparse-prior guided diffusion framework for point cloud compression, designed to reduce latent redundancy while preserving high-fidelity geometry under aggressive bitrates.

Abstract

While conventional lossy compression methods predominantly depend on autoencoders to map point clouds into latent representations, they often neglect the intrinsic redundancy within these latent points. To address this limitation, DiffCom presents a diffusion-based architecture steered by sparse priors, designed to minimize latent redundancy while securing superior reconstruction fidelity, particularly in low-bitrate scenarios.

A key feature of the framework is an efficient dual-density data flow that alleviates the stringent size constraints imposed on latent points. By integrating a Probabilistic Attention-based Conditional Denoiser (PACD), the method encapsulates critical reconstruction details within sparse priors, which are hierarchically decoupled into intra- and inter-point components. Comprehensive experiments on ShapeNet and standard MPEG PCC datasets demonstrate superior rate-distortion trade-offs.

Pipeline

The pipeline employs a two-stage data flow. First, DiffCom extracts sparse priors through a sparse point encoder and context-aware entropy model. Then, during decompression, it starts from Gaussian noise and applies PACD conditioned on the sparse priors to reconstruct dense latent representations, which are decoded into the final point cloud.

Comparison between conventional compression and DiffCom — **Figure 1.** (a) Conventional compressors code the latent representation of points directly via a naive context-free entropy model. (b) Our method codes the sparse priors via a context-aware entropy model. It employs a two-stage data flow to compress the points further into decoupled sparse priors. It incorporates a probabilistic attention-based conditional denoiser (PACD) model to denoise the latent representations conditioned on the sparse priors.

DiffCom framework overview — **Figure 2.** Overview of our decoupled sparse priors guided diffusion compression model for point cloud (DiffCom). Instead of directly encoding the input point cloud into latent points and features, we utilize a sparse point encoder to extract a sparser representation. The sparser representations are decoupled into sparse points and inter-point local distributions. During decompression, we begin with Gaussian noise and apply a Probabilistic attention-based conditional denoiser (PACD) conditioned on the sparse priors, reconstructing the latent representation. These reconstructed latents are then decoded to produce a high-quality point cloud.

Qualitative Results

DiffCom preserves more complete object structures at matched bitrates on ShapeNet and generalizes to dense MPEG point clouds.

ShapeNet qualitative comparison — Qualitative results on ShapeNet at two selected rates.

MPEG qualitative comparison — Qualitative comparison on longdress_vox10_1300 from the MPEG dataset.

Datasets

DiffCom is trained on ShapeNet and evaluated on object-level ShapeNet point clouds as well as standard MPEG PCC test samples from 8iVFB and Owlii.

ShapeNet training samples — Training samples from ShapeNet.

MPEG test samples — Test samples from 8iVFB and Owlii datasets.

Rate-Distortion Results

MPEG rate-distortion curve — Rate-distortion performance on the MPEG dataset.

ShapeNet rate-distortion curve — Rate-distortion performance on the ShapeNet dataset.

ShapeNet ablation on denoising steps — Ablation of denoising steps on ShapeNet.

Dataset	PSNR Gain	Rate Saving
MPEG average vs. G-PCC	+14.62 dB	98.46%
ShapeNet vs. Depoco	+6.89 dB	89.31%

Code

Code will be released when available.

Citation

@article{zhang2026diffcom,
  title={DiffCom: Decoupled Sparse Priors Guided Diffusion Compression for Point Clouds},
  author={Zhang, Xiaoge and Wu, Zijie and Feng, Mingtao and Nasim, Mehwish and Anwar, Saeed and Mian, Ajmal},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2026},
  publisher={IEEE}
}