Research map / Vision-language and multimodal / Multimodal LLMs
Exploring Spatial Intelligence from a Generative Perspective
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.
Introduces GSI-Bench (GSI-Real and GSI-Syn) to measure generative spatial intelligence via spatially grounded image editing, and shows fine-tuning BAGEL on simulator-generated GSI-Syn data improves both editing and downstream spatial understanding.
Problem
Spatial intelligence in MLLMs is assessed mainly from an understanding perspective; it is unclear whether generative/unified models can respect and manipulate 3D spatial constraints during image generation, and whether this can be measured and improved.
Approach
Operationalizes generative spatial intelligence as spatially grounded image editing with quantitative 3D operations, builds a real (ScanNet++-based, 3D-prior-guided) and a simulator-based synthetic benchmark with four metrics (IC, SA, AC, EL), and shows that generative fine-tuning on simulated edits transfers to spatial understanding.
Key results
- BAGEL + GSI-Syn improves GSI-Real average from 28.46 to 36.28 (+7.83) over BAGEL.
- On GSI-Syn-Table, BAGEL + GSI-Syn reaches 48.74 average vs 26.59 for BAGEL (+22.15).
- On OmniSpatial, BAGEL + GSI-Syn reaches 42.07 overall vs 41.55 for BAGEL; on SAT-Real overall 69.33 vs 65.33.
Surprising findings
- Fine-tuning purely on synthetic editing data improves spatial understanding benchmarks without any understanding or reasoning data.
- Emu3.5 (43.52 GSI-Real avg) outperforms closed-source Nano Banana and GPT-image on the real-world subset.
Limitations
- GSI-Syn-Room gains are modest due to scene complexity and spatial ambiguity.
- Complex Logic understanding decreased after GSI-Syn fine-tuning.
- GSI-Real relies on 3D grounding model (DetAny3D) and MLLM gating, so positional uncertainty remains.
Experimental setup
Datasets: GSI-Syn-Train ScanNet++ AI2-THOR MesaTask OmniSpatial SAT-Real
Benchmarks: GSI-Real GSI-Syn-Room GSI-Syn-Table OmniSpatial SAT-Real
Metrics: Instruction Compliance Spatial Accuracy Appearance Consistency Edit Locality
Baselines: Nano Banana GPT-image AnyEdit UniWorld Ultra Qwen-Image-Edit OmniGen2 Emu3.5 BAGEL
Builds on: BAGEL DetAny3D AI2-THOR MesaTask
Closest papers by meaning
- MMSI-Bench: A Benchmark for Multi-Image Spatial Intelligence (ICLR 2026 · similarity 0.91)
- Scaling Spatial Intelligence with Multimodal Foundation Models (CVPR 2026 · similarity 0.91)
- GGBench: A Geometric Generative Reasoning Benchmark for Unified Multimodal Models (CVPR 2026 · similarity 0.90)
- SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence (CVPR 2026 · similarity 0.90)
- SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- From Where Things Are to What They Are For: Benchmarking Spatial-Functional Intelligence in Multimodal LLMs (CVPR 2026 · similarity 0.89)
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.