Research map / Point clouds / Shape and surface completion

GeoComplete: Geometry-Aware Diffusion for Reference-Driven Image Completion

NeurIPS 2025 · Point clouds › Shape and surface completion · approach: Point cloud completion networks · 1 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 GeoComplete, a dual-branch Stable Diffusion inpainting model conditioned on VGGT-derived projected point clouds (with LangSAM dynamic-object filtering) and trained with target-aware masking to complete missing target regions from reference images.

Problem

Reference-driven image completion restores missing regions in a target view using other views of the same scene, but generative methods without geometric cues produce misaligned or implausible content when the target differs strongly from the references.

Approach

Condition a per-scene fine-tuned diffusion inpainting model on projected point clouds from VGGT (with dynamic objects removed by LangSAM) via a second branch joined by masked joint self-attention, and mask reference regions that are visible in references but absent from the target (target-aware masking) during training.

Key results

• On RealBench, GeoComplete reaches 17.32 PSNR and 0.578 SSIM versus 14.78 PSNR and 0.424 SSIM for RealFill, the strongest baseline.
• Ablation shows dual-branch geometry conditioning raises PSNR from 14.78 to 16.37, joint self-attention with mask to 16.85, and target-aware masking to 17.32.
• Under 75% point-cloud noise GeoComplete still reaches 16.90 PSNR, remaining above RealFill's 14.78 clean-input PSNR.

Measured effects reported in the paper

Surprising findings

• Reference-guided image editing models (OmniGen, Step1X-Edit, Bagel) often produce blank or meaningless outputs and fail on many scenes, so they were evaluated only on subsets.

Limitations

• Quality depends on VGGT geometry accuracy; imprecise point clouds limit exploitation of geometric cues.
• Performance degrades when reference images suffer from rain, haze, or low light.
• Per-scene 2,000-step fine-tuning is the computational bottleneck.

Experimental setup

Datasets: RealBench QualBench

Benchmarks: RealBench QualBench

Metrics: PSNR SSIM LPIPS DreamSim DINO CLIP user study score

Baselines: SD Inpaint Generative Fill Paint-by-Example TransFill RealFill OmniGen Step1X-Edit Bagel

Builds on: RealFill Stable Diffusion 2 Inpainting VGGT LangSAM

Closest papers by meaning

• Rethinking Multimodal Point Cloud Completion: A Completion-by-Correction Perspective (AAAI 2026 · similarity 0.89)
• Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.87)
• PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors (CVPR 2025 · similarity 0.87)
• Context and Geometry Aware Voxel Transformer for Semantic Scene Completion (NeurIPS 2024 · similarity 0.86)
• Point-based Instance Completion with Scene Constraints (ICLR 2025 · similarity 0.86)
• DeCoTR: Enhancing Depth Completion with 2D and 3D Attentions (CVPR 2024 · similarity 0.86)