wafer-bob/ASASR

What it solves

ASASR addresses the problem of "faithful" image super-resolution (SR), where the goal is to upscale low-resolution images while maintaining structural fidelity and spectral consistency without introducing the artificial artifacts common in many generative SR methods.

How it works

Built on the FLUX.1-dev backbone, ASASR uses a dual-LoRA inference pipeline (a base SR prior and a DPO refinement) to reconstruct images. It introduces three key technical innovations:

1. Colored-Noise Flow: Replaces standard Gaussian noise with a spectrally shaped kernel that aligns with the natural image manifold.
2. Sobolev-Induced Geometry (S-DPO): Uses a Riemannian metric and frequency-weighted Direct Preference Optimization (DPO) to better preserve high-frequency structures.
3. Adversarial Manifold Guidance (AMG): Employs an adversary network trained to mimic baseline SR artifacts, which then generates "hard negatives" to guide the DPO process toward more faithful reconstructions.

Who it’s for

Researchers and developers working on image restoration, super-resolution, and generative AI who need high-fidelity upscaling that avoids common generative hallucinations or artifacts.

Highlights

• SOTA Performance: Ranked as state-of-the-art on DIV2K-Val (Real-ESRGAN ×4) and #3 on RealSR.
• Dual-LoRA Architecture: Combines a base SR LoRA with a DPO-refined LoRA for balanced reconstruction.
• Adversarial Training: Uses a specialized adversary network to synthesize realistic failure cases for preference optimization.
• Comprehensive Evaluation: Includes a full suite of metrics including PSNR, SSIM, LPIPS, and no-reference quality assessments (MANIQA, MUSIQ, CLIPIQA+).