Research map / Vision-language and multimodal / Multimodal LLMs

GeoDiT: A Diffusion-based Vision-Language Model for Geospatial Understanding

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 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 GeoDiT, a masked-diffusion vision-language model built on LLaDA-8B with a SigLIP-2 encoder that generates remote-sensing text and boxes via parallel iterative refinement, and reports SOTA on captioning, grounding, detection, VQA and classification.

Problem

Autoregressive remote-sensing VLMs generate sequentially, which conflicts with the parallel, object-centric structure of geospatial understanding and causes failures in captioning and multi-object detection.

Approach

Reframes geospatial text generation as parallel mask-and-predict diffusion with low-confidence remasking, using a SigLIP-2 visual encoder feeding a LLaDA-8B bidirectional transformer, trained in two stages on SkyScript and MMRS-1M optical data.

Key results

• GeoDiT reaches CIDEr 135.6 on RSICD, a 17.6% relative gain over EarthDial (per paper).
• On VRSBench, GeoDiT achieves VG Acc@0.5 of 63.7 and DET mAP@0.5 of 24.9, above the strongest baselines in each category.
• Low-confidence remasking beats random remasking on RSICD CIDEr (135.6 vs 121.8) and DIOR-RSVG mAP (20.8 vs 15.5).

Surprising findings

• Diffusion-based general-domain VLMs (LLaDA-V, LaViDA, MMaDA) score 0.0 on all visual grounding and detection benchmarks in the reported table.

Limitations

• Gains are reported mainly on structured, object-centric tasks; narrative captioning gains are smaller (BLEU-4 modest).
• Inference speed and computational cost compared with autoregressive baselines are not quantified.
• Baselines like GPT-4V and Claude-4 are evaluated with unspecified prompting; comparison fairness across training data is unclear.

Experimental setup

Datasets: SkyScript MMRS-1M UCM-Captions RSICD RSITMD Sydney-Captions NWPU-Captions FloodNet RSVQA LR RSIVQA CRSVQA NWPU-RESISC45

Benchmarks: UCM-Captions RSICD NWPU-Captions Sydney-Captions DIOR-RSVG VRSBench AVVG RSVG RSVQA-LR RSVQA-HR AID WHU-RS19

Metrics: BLEU-4 METEOR CIDEr Acc@0.5 mAP@0.5 accuracy

Baselines: GPT-4V Claude-4 LLaDA-V LaViDA MMaDA LLaVA-1.5 Qwen2.5-VL GeoChat VHM EarthDial

Builds on: LLaDA DiT SigLIP-2 LLaVA-1.5

Closest papers by meaning

• Textual Supervision Enhances Geospatial Representations in Vision-Language Models (ICML 2026 · similarity 0.90)
• GeoViS: Geospatially Rewarded Visual Search for Remote Sensing Visual Grounding (CVPR 2026 · similarity 0.89)
• GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.88)
• GeoDiffusion: Text-Prompted Geometric Control for Object Detection Data Generation (ICLR 2024 · similarity 0.87)
• Stitch and Tell: A Structured Data Augmentation Method for Spatial Understanding (NeurIPS 2025 · similarity 0.86)
• TimeSpot: Benchmarking Geo-Temporal Understanding in Vision–Language Models in Real-World Settings (ICML 2026 · similarity 0.86)