FUSAR-GPT : A Spatiotemporal Feature-Embedded and Two-Stage Decoupled Visual Language Model for SAR Imagery
Paper • 2602.19190 • Published
FUSAR-GPT
Model weights for FUSAR-GPT, accepted by CVPR 2026.
| Stage | Task | Directory |
|---|---|---|
| Stage 1 | Alignment | stage1 |
| Stage 2 | Spatial localization (Task 1) | stage2/localization |
| Stage 2 | Object counting (Task 2) | stage2/counting |
| Stage 2 | Object classification (Task 3) | stage2/classification |
| Stage 2 | Object detection (Task 4) | stage2/detection |
Each released model directory contains its task adapter, TLM weights, and processor/tokenizer files. The Qwen2.5-VL-7B-Instruct backbone is downloaded separately.
Download
Download only the detection model:
hf download kkcyy/FUSAR-GPT --include "stage2/detection/*" --local-dir ./checkpoints
The files are saved in ./checkpoints/stage2/detection/. Replace detection with localization, counting, or classification to select another task.
Download the Stage 1 alignment model:
hf download kkcyy/FUSAR-GPT --include "stage1/*" --local-dir ./checkpoints
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for kkcyy/FUSAR-GPT
Base model
Qwen/Qwen2.5-VL-7B-Instruct