SHARP: Spectrum-aware Highly-dynamic Adaptation for Resolution Promotion in Remote Sensing Synthesis

Dimensions overall score 8.0

• Training-free resolution promotion via Rotary Position Embedding (RoPE) rescaling offers a practical remedy for the prohibitive cost of training remote sensing synthesis models at large resolutions.technical 85%
• Existing training-free resolution promotion methods apply a static positional scaling rule throughout denoising, which is harmful for RS imagery due to its dense medium- and high-frequency energy.limitation 90%
• The proposed SHARP method introduces a rational fractional time schedule k_rs(t) into RoPE to apply strong positional promotion early and relax it progressively, aligning with the frequency-progressive nature of diffusion denoising.method 90%
• SHARP's resolution-agnostic formulation enables robust multi-scale generation from a single set of hyperparameters.technical 85%
• Extensive experiments across six square and rectangular resolutions show SHARP consistently outperforms all training-free baselines on CLIP Score, Aesthetic Score, and HPSv2.result 90%
• The performance margins of SHARP over baselines widen at more aggressive extrapolation factors.result 85%
• SHARP achieves its superior performance with negligible computational overhead.technical 85%
• Remote sensing synthesis lags behind general text-to-image generation due to the absence of a domain-specialized DiT prior and the prohibitive cost of training at large resolutions.limitation 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card