AsySPLIT: Efficient 3DGS Scene Modeling

HKUST · hf · 2026-07-17

AsySPLAT proposes an efficient, heterogeneous 3D Gaussian Splatting architecture for **long-sequence scene modeling** that decouples geometry and appearance modeling. The paper's premise is that while existing general 3DGS methods improve long-sequence novel view synthesis (NVS), they suffer from massive computational redundancy. The authors argue that high-quality NVS doesn't necessarily require extremely precise geometry, and appearance learning is generally easier than geometry recovery. Thus, they designed a **task-aware asymmetric architecture**: - **Geometry branch**: Processes coarse-grained tokens, handles multi-view reconstruction, and accounts for a larger proportion of parameters. - **Appearance branch**: Processes fine-grained tokens, capturing details with fewer parameters. - **Bidirectional connections** allow the two branches to exchange information and assist each other's tasks. Experimentally, this design significantly reduces computational redundancy and improves parameter efficiency: - On **32-view 960P** inputs, the model achieves results comparable to optimization-based methods while delivering nearly an **800x speedup**. - Compared to the current strongest general zero-shot methods, it achieves better performance with **fewer parameters** and lower training/inference overhead.

Original post →

More from Research

• Draft paper uses Markov-chain eigenfunctions to build partitions and speed up sampling — michaelchchoi · 2026-07-21
• Autoresearch proposes packaging ML runs as studies with questions, analysis, and code diffs — morgymcg · 2026-07-21
• GitHub repo adds lightweight ternary QAT for Prism-ML Bonsai models — terminoid_ · 2026-07-21
• Qdrant co-hosts a Munich meetup on search, retrieval, and agentic RAG on July 23 — qdrant_engine · 2026-07-21
• GigaChat Audio targets long-form audio grounding with timestamps across 120-minute inputs — ai-sage · 2026-07-21
• Paper models Transformer components as stochastic geometry and tests five architectures — Zhihua Liang · 2026-07-21

Research channel →