Search papers, labs, and topics across Lattice.
This paper introduces PixelIR, a novel framework for real-world image super-resolution that decouples fidelity and perceptual detail synthesis to enhance efficiency. By utilizing pixel-space image-residual flow matching, PixelIR first reconstructs a faithful image from degraded observations and then synthesizes perceptual details from noise without compromising the overall restoration. Extensive evaluations show that PixelIR outperforms existing methods in PSNR, SSIM, and LPIPS metrics while maintaining a lightweight architecture with low latency.
Achieving state-of-the-art performance in real-world image super-resolution while maintaining a rapid single-step processing time and minimal computational overhead.
Real-world image super-resolution (Real-ISR) aims to preserve structures supported by the degraded observation while reconstructing perceptually realistic details. However, existing Real-ISR methods largely optimize fidelity and perceptual quality within a shared network, causing the two objectives to interfere throughout training and making their balance difficult to control. Recent one-step methods reduce sampling steps, yet often inherit both this coupled optimization behavior and the expensive high-resolution backbone of their multi-step predecessors. We argue that efficient Real-ISR requires not only a shorter sampling trajectory, but also specialized modeling of faithful reconstruction and perceptual detail synthesis. Based on this insight, we propose PixelIR, a fidelity-perception decoupling framework built upon pixel-space image-residual flow matching. PixelIR first learns an image flow that maps the degraded observation to a faithful reconstruction. Then, a residual flow synthesizes the missing perceptual details from noise without repeatedly relearning or overwriting the complete restoration solution. We further distill the teacher into a deployment-oriented one-step student within a coarse-to-fine pyramid architecture. Extensive experiments show that PixelIR achieves leading PSNR, SSIM, and LPIPS on both RealSR and DRealSR. The final model completes pixel-space restoration in a single evaluation with only 32.9M parameters, 89.7G MACs, and 8.5ms latency, demonstrating a strong practical fidelity-perception-efficiency balance.
