Stellaria . Research

Multi-Resolution
Enhancement Model.

Redefining satellite imagery with AI super-resolution. From 30 cm down to 3 cm class detail, turning coarse object awareness into analytics-ready insight for defense, energy, and smart-city workflows.

View Results Benchmark Metrics
Hallucination
< 1%
Analytics lift
+15 to 25%
Inference
0.49 s / tile
Scale
Up to 10x

01 . Abstract

Operational super-resolution for real satellite data.

Meruem, short for Multi-scale Enhanced Resolution using Unified Embedding Models, is Stellaria's next-generation system for satellite image super-resolution. It is built for real-world multi-sensor imagery, preserves geospatial integrity, and improves downstream analytics such as detection, segmentation, and monitoring. Unlike traditional super-resolution pipelines that focus on visual sharpness, Meruem is designed for operational deployment.

02 . Overview

Resolution is the bottleneck. Meruem moves it.

Three resolution thresholds map to three operating modes, from object awareness to analytics-ready detail.

Why super-resolution matters

From object awareness to operational insight.

  • At 30 cm, you see objects.
  • At 10 cm, you understand structure.
  • At 3 cm class detail, you unlock operational insight.
  • Resolution is a core bottleneck across GEOINT workflows.

What Meruem is designed to do

Built for analytics, not just appearance.

  • Enhance spatial resolution by up to 10x.
  • Preserve geospatial accuracy.
  • Support real satellite data and multi-sensor inputs.
  • Improve downstream analytics, not image appearance alone.
30 cm
Object awareness
10 cm
Structural understanding
3 cm class
Operational detail
Up to 10x
Spatial enhancement

03 . Method

How it works.

A three-stage pipeline that preserves scene integrity from input to super-resolved output.

  1. 01 . Input

    Real satellite imagery

    Real satellite tiles with sensor variation, noise, and compression artifacts.

  2. 02 . Encode & fuse

    Multi-resolution encoding

    Cross-scale feature extraction preserves structure, while a unified embedding combines context for enhancement and reliability control.

  3. 03 . Output

    Super-resolved imagery

    Analytics-oriented output for detection, segmentation, and monitoring, with geospatial integrity preserved end to end.

04 . Qualitative Results

Drag to compare input against Meruem's output.

Real satellite crops at native scale. Drag each handle to swipe between the low-resolution input and Meruem's super-resolved output.

Before After

Crop . 01

Vehicle and lot

Recovers vehicle outline, vegetation structure, and ground texture.

Before After

Crop . 02

Road intersection

Sharpens lane markings, crosswalks, and individual vehicles.

05 . Quantitative Results

Strong fidelity and perceptual scores.

Meruem maintains high pixel-level fidelity while excelling on perceptual and no-reference quality metrics, with faithful reconstruction paired with visually convincing detail.

< 1% Hallucination rate Scene-integrity preserved
+15 to 25% Analytics improvement Detection & segmentation lift
0.49s Inference per tile Near real-time on GPU
Metric Direction What it measures Value
PSNR Higher is better Pixel-level accuracy 37.12 dB
SSIM Higher is better Structural similarity 0.953
LPIPS Lower is better Perceptual similarity 0.134
NIQE Lower is better No-reference quality 4.24
MUSIQ Higher is better Multi-scale perception 73.5
CLIPIQA Higher is better Semantic quality 0.788

06 . Impact & Reliability

Designed for real-world workflows.

Downstream analytics

Lift across detection and segmentation.

  • +15% to +25% improvement in object detection and segmentation.
  • Improved edge clarity and feature extraction.
  • Maintains geolocation accuracy for real use cases.

Reliability profile

Faithful by design.

  • Hallucination rate below 1%.
  • Sensor-aware optimization adds up to +12% performance.
  • Robust to atmospheric distortion and compression artifacts.

07 . Product Tiers

Three tiers, one model family.

Meruem Mini

2x

  • Lightweight, fast enhancement.
  • Ideal for visualization and preprocessing.
  • Planned open-source GitHub release.

Meruem 4x

4x

  • High-quality enhancement for detection and monitoring.
  • Optimized for production analytics.
  • Available through the Stella Platform.

Meruem Ultra

8 to 10x

  • High-detail enhancement with advanced refinement.
  • Designed for secure or enterprise environments.
  • Available as an enterprise solution.

08 . Performance & Deployment

Operational from day one.

Processing speed Near real-time, GPU optimized 0.49 s / tile
Hardware NVIDIA A100 / H100 / RTX class Inference-ready
Throughput Batch and streaming Scalable
Deployment On-prem or secure environments Operationally compatible
Integration GEOINT platform compatibility e.g. BAE Systems GXP

09 . Early Access

Available now, and on the open-source roadmap.

Available now

Operational deployments via Stellaria.

  • Meruem 4x through the Stella Platform.
  • Meruem Ultra for enterprise deployment.

Open-source track

For researchers and developers.

  • Meruem Mini will be released on GitHub.
  • Designed for rapid prototyping and research.
Request Early Access Contact Stellaria