11/ These improvements were validated on two large-scale benchmarks (DL3DV-10K and RealEstate10K) and three well-known foundation models (DepthSplat, MVSplat and pixelSplat).

10/ 🔍Enhanced Efficiency: ZPressor effectively prevents the linear growth of inference time and memory consumption that typically plagues existing methods with increasing input views. Specifically, for 36 input views on DepthSplat, ZPressor reduces memory footprint by 80% and inference time by 70%.

9/ 🔍Consistent Performance Improvement: Under moderate input view settings, ZPressor consistently improves the performance of these models. For instance, with 36 input views, ZPressor boosts the PSNR on DepthSplat by 4.65 dB

8/ 🔍Scalability to Dense Views: ZPressor enables existing models to scale to over 100 input views at 480P resolution, even on GPUs with limited memory (e.g., 80GB GPU).

7/ Experimental Results: Enhanced Performance and Robustness
Our experiments demonstrate the significant benefits of integrating ZPressor into state-of-the-art feed-forward 3DGS models:

6/ This plug-and-play module effectively enhances the capacity of existing feed-forward 3DGS models to handle a larger number of input views.

5/ 🔍Information Fusion: Concretely, ZPressor enables this compression by partitioning input views into "anchor" and "support" sets. It then uses a lightweight architecture-agnostic module to compress information from the support views into the anchor views, forming the compact latent state.

4/ 🔍Redundancy Discarding: It intelligently retains essential scene information while actively discarding redundant data present across multiple views.

3/ 🔍Information Bottleneck Principle: It analyzes feed-forward 3DGS frameworks through the lens of the Information Bottleneck principle. This allows ZPressor to efficiently compress multi-view inputs into a compact latent state.

2/ The Solution: ZPressor – Bottleneck-Aware Compression
To address these scalability issues, we introduce ZPressor, a lightweight and architecture-agnostic module. ZPressor tackles the problem by:

1/ The Problem: Scalability Constraints in Feed-Forward 3DGS

Essentially, existing feed-forward 3DGS frameworks struggle to scale efficiently when dealing with a high density of input views.

Existing feed-forward 3DGS models struggle with dense views, facing performance drops & massive redundancy. ZPressor leverages Information Bottleneck Theory to compress multi-view features, significantly boosting scalability and reconstruction quality for robust dense-view synthesis.