Research map / Architectures and efficiency / Attention and transformers
Efficient-LVSM: Faster, Cheaper, and Better Large View Synthesis Model via Decoupled Co-Refinement Attention
Xiaosong Jia, Yihang Sun, Junqi You, Songbur Wong, Zichen Zou, Junchi Yan, Zuxuan Wu, Yu-Gang Jiang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add self-then-cross attention in decoder (ablation) | Efficient-LVSM 6+6 · RealEstate10K | PSNR (dB): 24.18 -> 24.97 | Improved |
| add encoder-decoder co-refinement (ablation) | Efficient-LVSM 6+6 · RealEstate10K | PSNR (dB): 24.97 -> 26.25 | Improved |
| add REPA distillation (ablation) | Efficient-LVSM 6+6 · RealEstate10K | PSNR (dB): 26.02 -> 26.81 | Improved |
| add REPA distillation (ablation) | LVSM Dec-Only · RealEstate10K | PSNR (dB): 25.52 -> 25.68 | Improved |
| use Smooth L1 REPA loss instead of L2 or cosine (ablation) | Efficient-LVSM REPA · RealEstate10K | PSNR (dB): 26.39 -> 26.81 (L2); 26.30 -> 26.81 (Cosine) | Improved |
| distill both input and target tokens instead of only one (ablation) | Efficient-LVSM REPA · RealEstate10K | PSNR (dB): 26.35 -> 26.60 | Improved |
| use DINOv3 layer 8 instead of layer 10 or 12 (ablation) | Efficient-LVSM REPA · RealEstate10K | PSNR (dB): 26.28 -> 26.60 (layer 10); 26.11 -> 26.60 (layer 12) | Improved |
| increase model from 3+3 to 6+6 to 12+12 layers (ablation) | Efficient-LVSM · RealEstate10K | PSNR (dB): 26.43 -> 27.77 -> 28.32 | Improved |
| add attention masks to LVSM (ablation) | LVSM · RealEstate10K | PSNR (dB): 25.24 -> 24.13 | Hurt |
| add MMDiT-style separate projections to LVSM (ablation) | LVSM · RealEstate10K | PSNR (dB): 25.24 -> 24.37 | Hurt |
| use decoupled dual-stream design instead of LVSM Dec-Only | Efficient-LVSM · RealEstate10K inference latency | latency (ms): 109.37 ms -> 24.78 ms | Improved |
| use decoupled dual-stream design instead of LVSM Dec-Only (512 res) | Efficient-LVSM · RealEstate10K (res-512) | PSNR (dB): 29.53 -> 29.86 | Improved |
Closest papers by meaning
- LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias (ICLR 2025 · similarity 0.90)
- A-VL: Adaptive Attention for Large Vision-Language Models (AAAI 2025 · similarity 0.86)
- Resolving Representation Ambiguity in Feedforward Novel View Synthesis Transformer via Semantic-Spatial Decoupling (NeurIPS 2026 · similarity 0.86)
- NVComposer: Boosting Generative Novel View Synthesis with Multiple Sparse and Unposed Images (CVPR 2025 · similarity 0.85)
- SmallKV: Small Model Assisted Compensation of KV Cache Compression for Efficient LLM Inference (NeurIPS 2025 · similarity 0.85)
- Dynamic View Synthesis as an Inverse Problem (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
