ZJU's InfiniSplat v2 Cuts Feed-Forward 3DGS Perceptual Error by 32.9% (118 HF Upvotes), While Technion's H-DDCM Enables Single-GPU Streaming Generative Video Compression

Core Background & Industry Pain Points

In real-time 3D scene reconstruction and ultra-low-bitrate visual communication, feed-forward 3D Gaussian Splatting (3DGS) and training-free diffusion compression face critical scaling bottlenecks: (1) while feed-forward 3DGS models reconstruct 3D scenes directly from sparse input images, learning Gaussian appearance attributes atop predicted geometry degrades sharply under large viewpoint changes; and (2) codebook-based training-free generative compression methods (such as Turbo-DDCM) rely on dense Gaussian codebooks and combinatorial bitstream encoding that scale poorly in memory and compute with the high-dimensional latent spaces of modern image and video generators.

Architectural Highlights & Core Mechanisms

Two new works—Zhejiang University's InfiniSplat v2 (arXiv:2610.20341, 118 upvotes on Hugging Face) and Technion's H-DDCM (arXiv:2610.18125)—introduce principled architectural solutions:

• ZJU's InfiniSplat v2: Decoupled Geometry/Attribute Learning & Large-Angle Supervision: Decouples geometry and Gaussian attribute prediction by freezing the geometry branch, aligns predicted Gaussians using ground-truth camera poses during training, and enforces large-angle target view supervision—demonstrating that opacity learned solely from image reconstruction supervision naturally suppresses geometrically inconsistent points.
• Technion's H-DDCM: Implicit Randomized Hadamard Dictionaries & Streaming KV-Prefix Video Compression: Replaces dense Gaussian codebooks with implicit randomized Hadamard dictionaries via the fast Walsh-Hadamard transform (FWHT), achieving $O(d \log d)$ compute and $O(d)$ memory without materializing full dictionaries. Combined with a table-free lexicographic coding scheme and periodic re-encoding of reconstructed frames into a fresh KV-cache prefix for autoregressive video flow models, H-DDCM preserves temporal context across segments at zero extra transmitted bits.

Authoritative Benchmarks & Empirical Comparisons

• InfiniSplat v2 Sets SOTA Across 6 Datasets (arXiv:2610.20341): Evaluated across DL3DV, ETH3D, NRGBD, Mip-NeRF 360, RE10K, and DTU ($N=4,8,16$ views), a single trained InfiniSplat v2 model achieves state-of-the-art novel view synthesis in both pose-free (mean LPIPS 0.3026, DISTS 0.1341 vs. WorldMirror 2.0 at 0.3757 / 0.1935) and pose-conditioned settings—reducing mean LPIPS by 25.8% (0.2472) and mean DISTS by 32.9% (0.1163) compared with the strongest baselines.
• H-DDCM Enables Single-GPU Near-Real-Time Generative Video Compression (arXiv:2610.18125): Across image and video benchmarks, H-DDCM matches dense-codebook rate-distortion-perception quality while slashing codebook memory to $O(d)$, enabling long-form, near-real-time generative video compression on a single GPU.

Developer Practical Guide & Quickstart

3D vision engineers can access InfiniSplat v2 code and weights at https://github.com/zju3dv/InfiniSplat (https://zju3dv.github.io/InfiniSplat2/), while media compression developers can explore H-DDCM's Hadamard dictionary and KV-prefix streaming implementation at https://h-ddcm.github.io/.