---
library_name: pytorch
pipeline_tag: image-to-image
tags:
  - super-resolution
  - remote-sensing
  - flow-matching
  - gaussian-splatting
  - continuous-scale
---

# FlowGS pretrained models

Pretrained models for **Flow-based Gaussian Splatting for Continuous-Scale
Remote Sensing Image Super-Resolution**.

- Paper: https://doi.org/10.1109/LGRS.2026.3729993
- arXiv: https://arxiv.org/abs/2605.22147
- Code: https://github.com/hanlinwu/FlowGS

## Files

| File | Purpose | Size | SHA-256 |
| --- | --- | ---: | --- |
| `first_stage_gs_res_v1.ckpt` | Continuous-scale Gaussian renderer required by the final config | 189 MiB | `68990a8884f2f27f6912664333698409b4047a1352e31317c49b5deaa1620f1b` |
| `flowgs.ckpt` | Final one-step FlowGS checkpoint | 1.3 GiB | `5874d635d1a86a6b33c44dc43c518008f094760b9e1067a54b9d610dff81f174` |

Both checkpoints are required for evaluation.

## Download

```bash
hf download hlwu/FlowGS --include "*.ckpt" --local-dir checkpoints
```

## Evaluation

Place the files under `checkpoints/` in the code repository, then run:

```bash
python test.py \
  --config configs/second_stage_fm_shortcut_1.yaml \
  --checkpoint checkpoints/flowgs.ckpt \
  --datasets AID,DOTA_512 \
  --scales 2.6,3.4,6.0,10.0 \
  --gpu 0
```

See the code repository README for environment setup, dataset preparation,
training stages, and DIOR evaluation.

## Citation

```bibtex
@article{mo2026flow,
  title   = {Flow-Based Gaussian Splatting for Continuous-Scale Remote Sensing Image Superresolution},
  author  = {Mo, Jiangwei and Lu, Xi and Wu, Hanlin},
  journal = {IEEE Geoscience and Remote Sensing Letters},
  volume  = {23},
  pages   = {6019205--6019205},
  year    = {2026},
  doi     = {10.1109/LGRS.2026.3729993}
}
```