Pixel to Gaussian: Ultra-Fast Continuous Super-Resolution with 2D Gaussian Modeling

2025-03-09

AI Summary

Arbitrary-scale super-resolution (ASSR) suffers from low efficiency and limited reconstruction fidelity in implicit neural representation (INR)-based approaches. To address this, we propose Pixel-to-Gaussian, a novel paradigm that replaces coordinate-based implicit neural representations with explicit continuous signal modeling via 2D Gaussian splatting. Specifically, our method directly synthesizes a statistical-feature-driven Gaussian parameter field from a low-resolution (LR) image; once constructed, this field enables millisecond-level (1 ms per scale) arbitrary-scale rendering without re-inference. By decoupling resolution from fixed-scale constraints and pixel-wise decoding, our approach achieves state-of-the-art (SOTA) reconstruction quality across multi-scale benchmarks while delivering unprecedented inference efficiency—significantly outperforming existing INR and CNN-based methods in both accuracy and speed.

Technology Category

• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Machine Learning: Representation Learning
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

Arbitrary-scale super-resolution (ASSR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs with arbitrary upsampling factors using a single model, addressing the limitations of traditional SR methods constrained to fixed-scale factors ( extit{e.g.}, $ imes$ 2). Recent advances leveraging implicit neural representation (INR) have achieved great progress by modeling coordinate-to-pixel mappings. However, the efficiency of these methods may suffer from repeated upsampling and decoding, while their reconstruction fidelity and quality are constrained by the intrinsic representational limitations of coordinate-based functions. To address these challenges, we propose a novel ContinuousSR framework with a Pixel-to-Gaussian paradigm, which explicitly reconstructs 2D continuous HR signals from LR images using Gaussian Splatting. This approach eliminates the need for time-consuming upsampling and decoding, enabling extremely fast arbitrary-scale super-resolution. Once the Gaussian field is built in a single pass, ContinuousSR can perform arbitrary-scale rendering in just 1ms per scale. Our method introduces several key innovations. Through statistical ana

Problem

• Reconstruct high-resolution images from low-resolution inputs.
• Enable arbitrary-scale super-resolution with a single model.
• Improve efficiency and fidelity using Gaussian Splatting.

Innovation

• Pixel-to-Gaussian paradigm for continuous super-resolution
• 2D Gaussian modeling eliminates upsampling and decoding
• Arbitrary-scale rendering in 1ms per scale