Research map / Image generation / Image restoration
COSMIC: Compress Satellite Image Efficiently via Diffusion Compensation
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 COSMIC, a satellite image compressor with a lightweight on-board encoder and a ground-side conditional latent diffusion model that compensates lost detail using satellite sensor metadata as conditions.
Problem
Satellite earth-observation images exceed satellite-to-ground transmission limits, but sophisticated learned compression encoders are too computationally heavy for onboard embedded GPUs with limited power.
Approach
Use a lightweight encoder (depthwise-conv LCB blocks plus cheap global attention via 1D convolutions) on the satellite, and on the ground a stable-diffusion-based compensation model conditioned on the latent code and sensor metadata (coordinates, timestamp, GSD, cloud cover, sun/camera angles) to restore lost details.
Key results
- On the tile test set, COSMIC uses 4.9 GFLOPs on-satellite versus 11.87-26.4 GFLOPs for baselines (about 2.6-5x fewer) while achieving the best LPIPS (0.0359) and FID (13.50).
- COSMIC outperforms baselines on perceptual metrics (LPIPS, FID) across bitrates, but has lower MS-SSIM at low bpp.
- Random-seed variance is small (e.g., PSNR std about 0.02-0.06 dB across 5 seeds).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use diffusion compensation during decoding instead of removing the diffusion module (ablation) | COSMIC · fMoW test set | LPIPS | Helped |
| Use diffusion compensation during decoding instead of removing the diffusion module (ablation) | COSMIC · fMoW test set | perceptual metrics | Helped |
| Keep the CAM module in the encoder rather than removing it (ablation) | COSMIC · fMoW test set | rate-distortion performance | Helped |
| Use metadata encoder for diffusion generation rather than image coding alone (ablation) | COSMIC · fMoW test set | reconstruction quality | Helped |
| Increase the number of reverse diffusion denoising steps (ablation) | COSMIC · fMoW test set | perceptual metrics | Helped |
| Increase the number of reverse diffusion denoising steps (ablation) | COSMIC · fMoW test set | PSNR | Hurt |
| Use COSMIC compression instead of the compared baselines at low bpp | COSMIC · fMoW and tile test sets | MS-SSIM | Mixed |
| Use COSMIC instead of the compared baselines across transmission rates | COSMIC · fMoW and tile test sets | LPIPS | Helped |
| Use COSMIC instead of the compared baselines across transmission rates | COSMIC · fMoW and tile test sets | FID | Helped |
| Use COSMIC instead of JPEG2000 for image classification after low-bitrate compression | COSMIC · Remote-sensing image classification (10, 15, and 20 classes) | accuracy change: -1.06% vs -4.21%; -0.70% vs -4.17%; -1.06% vs -3.16% | Helped |
Surprising findings
- At low bpp (under about 0.1), the diffusion compensation can generate non-existent textures, e.g., over ocean regions.
- Elic achieves higher PSNR and MS-SSIM on the tile set than COSMIC despite higher FLOPs.
Limitations
- Performance degrades at extremely low bpp (below about 0.1), where latent code lacks enough information to guide diffusion.
- Uses only a pretrained Stable Diffusion finetuned on natural-image priors, lacking satellite-specific prior knowledge.
- Diffusion may hallucinate textures that do not exist in the original image.
Experimental setup
Datasets: fMoW
Benchmarks: fMoW test set tile test set
Metrics: PSNR MS-SSIM LPIPS FID FLOPs bpp
Baselines: Elic Hific COLIC CDC HL_RS JPEG2000
Builds on: stable diffusion DDIM Elic CDC
Closest papers by meaning
- AstroCompress: A benchmark dataset for multi-purpose compression of astronomical data (ICLR 2025 · similarity 0.87)
- Towards Loss-Resilient Image Coding for Unstable Satellite Networks (AAAI 2025 · similarity 0.86)
- DiffusionSat: A Generative Foundation Model for Satellite Imagery (ICLR 2024 · similarity 0.85)
- Satellite to GroundScape - Large-scale Consistent Ground View Generation from Satellite Views (CVPR 2025 · similarity 0.85)
- Hazy Low-Quality Satellite Video Restoration Via Learning Optimal Joint Degradation Patterns and Continuous-Scale Super-Resolution Reconstruction (CVPR 2025 · similarity 0.83)
- SatSynth: Augmenting Image-Mask Pairs through Diffusion Models for Aerial Semantic Segmentation (CVPR 2024 · similarity 0.83)
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.