Image & Video

GeoDiff-SAR II enables controllable SAR generation with 3D-driven diffusion

New model uses 3D CAD models to precisely control azimuth, depression, and polarization in SAR images

Deep Dive

Researchers propose GeoDiff-SAR II, a foundation diffusion model for Synthetic Aperture Radar (SAR) image generation. It introduces a Geometric-Electromagnetic Conditioning Map (GECM) to decouple geometry from scattering, enabling control over azimuth angle, depression angle, and polarization. Training uses real sparse-azimuth data; inference renders from 3D CAD models via ControlNet and LoRA on a FLUX backbone. Achieves consistent improvements in image fidelity, physical consistency, and downstream Automatic Target Recognition (ATR) performance.

Key Points
  • Introduces Geometric-Electromagnetic Conditioning Map (GECM) to decouple geometry from electromagnetic scattering
  • Enables simultaneous control over azimuth angle, depression angle, and polarization for SAR generation
  • Uses ControlNet and LoRA on FLUX backbone, improving ATR performance by up to 20% on real datasets

Why It Matters

Enables physically consistent SAR generation for defense and remote sensing applications with precise control over imaging conditions.