CVPR Findings2026
GEODE: Geometry-Guided Discrete Diffusion for Open-Vocabulary 3D Scene Graph Generation
Changqun Feng, Wangxiandi Yin, Xin Hu, Lei Zhao, Dongyang Zhang, Tao He
Abstract
3D Scene Graph Generation (3DSGG) seeks to build a structured representation of a 3D scene by localizing objects and predicting their pairwise relations. Recent efforts push beyond the closed-set to open-vocabulary (OV) recognition by leveraging large vision-language models (VLMs). However, these pipelines inherit VLMs' weakness in spatial reasoning, yielding brittle predicate predictions under near-depth ambiguities and symmetric layouts. We introduce GEODE, a GEOmetry-guided DiscrEte diffusion framework that jointly denoises node geometry and edge predicates over tokenized scene graphs. Our key idea is to generate a globally geometry-constrained scene graph with a discrete graph diffusion process. Concretely, we vector-quantize per-object geometric attributes (center/size/orientation) and predicate semantics into compact codebooks, and perform categorical diffusion with an intrastep geometry to relation schedule: at each reverse step, geometry tokens are sampled first and then condition predicate updates, enforcing geometric consistency throughout generation. A Geometry-aware Graph Transformer tightly couples node-edge reasoning via relation-conditioned selfattention and an orientation-aware decomposition of edge updates, enabling precise spatial grounding without sacrificing semantic expressivity. On 3DSSG, GEODE sets a new state of the art in both closed-set and OV regimes, with particularly large gains on spatial-only evaluation under a strict top-1 criterion, e.g., with 19 point improvements over the SOTA RelationField.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- High-Resolution Image Synthesis with Latent Diffusion Models cites Denoising Diffusion Probabilistic Models
- High-Resolution Image Synthesis with Latent Diffusion Models cites Denoising Diffusion Implicit Models
- Denoising Diffusion Implicit Models cites Denoising Diffusion Probabilistic Models
Builds on47
- Learning Transferable Visual Models From Natural Language SupervisionAlec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh et al.ICML 2021 · 47,906 citations
- Denoising Diffusion Probabilistic ModelsJonathan Ho, Ajay Jain, Pieter AbbeelNeurIPS 2020 · 35,902 citations
- High-Resolution Image Synthesis with Latent Diffusion ModelsRobin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser et al.CVPR 2022 · 13,123 citations
- Denoising Diffusion Implicit ModelsJiaming Song, Chenlin Meng, Stefano ErmonICLR 2021 · 11,743 citations
- BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language ModelsJunnan Li, Dongxu Li, Silvio Savarese, Steven C. H. HoiICML 2023 · 7,873 citations
Related papers
- Can We Build Scene Graphs, Not Classify Them? FlowSG: Progressive Image-Conditioned Scene Graph Generation with Flow MatchingXin Hu, Ke Qin, Wen Yin, Yuan-Fang Li et al.CVPR 2026
- SPADE: Spatial-Aware Denoising Network for Open-Vocabulary Panoptic Scene Graph Generation with Long- and Local-Range Context ReasoningXin Hu, Ke Qin, Guiduo Duan, Ming Li et al.ICCV 2025 · 1 citation
- Open-World 3D Scene Graph Generation for Retrieval-Augmented ReasoningFei Yu, Quan Deng, Shengeng Tang, Yuehua Li et al.AAAI 2026 · 2 citations
- HSGG: Training-Free Hierarchical Scene Graph Generation with Geometry-Guided Relation Reasoningyunzhe Liu, Wenbiao Liu, Lihui Cen, Zhe Qu et al.ICML 2026
- From Pixels to Graphs: Open-Vocabulary Scene Graph Generation with Vision-Language ModelsRongjie Li, Songyang Zhang, Dahua Lin, Kai Chen et al.CVPR 2024
