GRAPE(Gaussian Rendering for Accelerated Pixel Enhancement) Brings Fast and Lightweight Arbitrary Super-Resolution

Jung In Jang¹ · Kyong Hwan Jin¹
¹Korea University, Republic of Korea
Paper Video Code

Abstract

GRAPE enables arbitrary-scale super-resolution (ASSR) by predicting 2D anisotropic Gaussian primitives on an image-space grid and efficiently rasterizing them into high-resolution outputs.

Instead of using heavy decoder architectures, GRAPE adopts a single-layer Gaussian prediction head combined with a GPU-friendly splatting renderer. This design enables fast, cache-efficient, and highly parallel inference at arbitrary output resolutions.

With only 1.56M parameters and 1.10 GB peak GPU memory, GRAPE achieves 69.33 FPS on Urban100 (985×798) while maintaining high reconstruction fidelity.

Method

GRAPE is an end-to-end differentiable pipeline:

  1. Encoder extracts low-resolution feature maps
  2. Feature Reshaping & Parameter Mapping converts features into a structured grid representation
  3. Gaussian Head (Point-wise Prediction) predicts anisotropic 2D Gaussian parameters:
    • RGB color
    • Rotation
    • Scale
    • Spatial offset
  4. 2D Rasterizer composites Gaussian primitives into the final high-resolution image in a single rendering pass

Method Overview

2D Anisotropic Gaussian Splatting

Oriented elliptical Gaussian footprints align naturally with edges and directional textures, improving structural fidelity and fine-detail reconstruction.

Below are visualizations of learned anisotropic parameters such as scale and rotation-related maps.

Anisotropic Visualization

Results

Qualitative Results (Urban100 Examples)

Qualitative Results

Quantitative Results

The following tables summarize quantitative comparisons across benchmark datasets.

Table 1
Table 2
Table 3

Runtime and Memory Efficiency

GRAPE reconstructs a Full-HD frame in 40.88 ms.
The encoder accounts for approximately 85.1% of total runtime, while the Gaussian head and rasterizer introduce minimal overhead.

Peak GPU memory usage remains below 2.12 GB, demonstrating strong efficiency for high-resolution inference.

Peak GPU Memory and Latency

Citation

@article{grape2025,
  title   = {GRAPE: Gaussian Rendering for Accelerated Pixel Enhancement},
  author  = {Jang, Jung In and Jin, Kyong Hwan},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2025}
}

Acknowledgements

This project page is adapted from the original project website template by: https://github.com/shunzh/project_website and customized for GRAPE.