Research map / Image generation / Fast diffusion sampling
Can Generative Geospatial Diffusion Models Excel as Discriminative Geospatial Foundation Models?
Yuru Jia, Valerio Marsocci, Ziyang Gong, Xue Yang, Maarten Vergauwen, Andrea Nascetti
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add learnable fusion (global weighted) instead of simple concatenation (ablation) | SatDiFuser · m-cashew-plantation | mIoU: 59.1 -> 66.5 | Improved |
| add learnable fusion (global weighted) instead of simple concatenation (ablation) | SatDiFuser · m-eurosat | top-1 accuracy: 94.5 -> 97.7 | Improved |
| add learnable fusion (global weighted) instead of simple concatenation (ablation) | SatDiFuser · m-so2sat | top-1 accuracy: 55.4 -> 59.3 | Improved |
| use multi-timestep multi-block features with fusion instead of a single timestep-1 self-attention block (ablation) | SatDiFuser · m-so2sat | top-1 accuracy: 53.6 -> 59.3 | Improved |
| use two experts activated (top-k 2) instead of one (ablation) | SatDiFuser MoE fusion · m-chesapeake-landcover | mIoU: 60.2 -> 71.6 | Improved |
| increase experts from 8 to 12 (ablation) | SatDiFuser MoE fusion · m-chesapeake-landcover | mIoU: 71.6 -> 68.3 | Hurt |
| reduce experts from 8 to 4 (ablation) | SatDiFuser MoE fusion · m-chesapeake-landcover | mIoU: 71.6 -> 69.8 | Hurt |
| use satellite-finetuned DiffusionSat instead of Stable Diffusion v2-1 backbone (ablation) | SatDiFuser Global fusion · m-chesapeake-landcover | mIoU: 65.8 -> 69.5 | Improved |
| use DiffusionSat instead of SD v2-1 backbone (ablation) | SatDiFuser Global fusion · m-cashew-plantation | mIoU: 63.7 -> 66.5 | Improved |
| add SatDiFuser global fusion versus best baseline with 10% labels | SatDiFuser vs CROMA · m-cashew-plantation (10% labels) | mIoU: 34.3 -> 39.6 | Improved |
| add SatDiFuser MoE fusion versus best pretrained model | SatDiFuser vs CROMA · m-bigearthnet | F1: 58.3 -> 65.8 | Improved |
| use sampling within first 20% of diffusion timesteps (ablation) | SatDiFuser features · m-eurosat and m-NeonTree | accuracy / mIoU | Improved |
Closest papers by meaning
- DiffusionSat: A Generative Foundation Model for Satellite Imagery (ICLR 2024 · similarity 0.92)
- SatSynth: Augmenting Image-Mask Pairs through Diffusion Models for Aerial Semantic Segmentation (CVPR 2024 · similarity 0.90)
- Parameter Efficient Self-Supervised Geospatial Domain Adaptation (CVPR 2024 · similarity 0.88)
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.88)
- MMM-RS: A Multi-modal, Multi-GSD, Multi-scene Remote Sensing Dataset and Benchmark for Text-to-Image Generation (NeurIPS 2024 · similarity 0.87)
- GeoDiT: A Diffusion-based Vision-Language Model for Geospatial Understanding (CVPR 2026 · similarity 0.86)
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.
