Research map / 3D vision / Gaussian splatting

MACRO: Training-free Multi-plane Attention for Closeup Render Optimization

NeurIPS 2026 · 3D vision › Gaussian splatting · approach: Regularization and primitive design for 3DGS · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MACRO, a training-free method that fixes scale mismatch in reference-conditioned diffusion enhancement of 3DGS close-up renders via depth-plane decomposition, scale-matched reference crops, and depth-masked attention, plus two new close-up NVS benchmarks.

Problem

Close-up novel view synthesis from 3D Gaussian splatting, where rendering zoomed far beyond training cameras degrades quality and reference-conditioned diffusion enhancers fail due to scale mismatch between close-up and reference views.

Approach

Decompose the close-up render into depth planes, crop and upsample reference images in image space to match each plane's scale before VAE encoding, and apply a depth-aware attention mask so each token attends only to scale-matched reference crops; no training required.

Key results

• On DL3DV-Closeup, MACRO reaches LPIPS 0.321, DreamSim 0.099, DINOv2 0.145, versus best baseline Difix at 0.337/0.123/0.203.
• On MobileClose-10, MACRO reaches PSNR 15.67, LPIPS 0.480, DreamSim 0.116, DINOv2 0.124, outperforming GSFixer and Difix on perceptual metrics.
• Ablation: removing the depth-aware mask modestly degrades results (e.g., DINOv2 0.145 to 0.161), and warping instead of cropping causes large degradation (LPIPS 0.391).

Surprising findings

• Progressive multi-step Difix3D variants often do not improve over single-step Difix and can degrade PSNR and perceptual metrics.
• Cosine similarity between VAE latents of resized images and resized latents drops to 0.70 at 4x scale, showing the VAE is not scale-equivariant.

Limitations

• Upsampled reference crops lose high-frequency detail, creating a blur-realism tradeoff.
• Number of depth planes P is fixed rather than adapted per view.
• Relies on 3DGS-rendered depth, which may be inaccurate where training views are sparse.

Experimental setup

Datasets: DL3DV-10K DL3DV-Closeup MobileClose-10

Benchmarks: DL3DV-Closeup MobileClose-10

Metrics: PSNR SSIM LPIPS DreamSim DINOv2 feature distance

Baselines: 3DGS Mip-Splatting SEVA GSFixer Difix Difix3D Difix3D+ PR-Difix3D PR-Difix3D+

Builds on: Difix 3D Gaussian Splatting SD-Turbo Difix3D

Closest papers by meaning

• SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis (CVPR 2025 · similarity 0.90)
• EditSplat: Multi-View Fusion and Attention-Guided Optimization for View-Consistent 3D Scene Editing with 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
• Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.89)
• FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.89)
• LP-3DGS: Learning to Prune 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
• GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)