GenAI Vision 2026

Image Enhancement & Super-Resolution Challenge

25-02-2026 09:30AM

00
Days
00
Hours
00
Mins
00
Secs
Explore Challenge

Background

In many real-world scenarios — surveillance systems, medical imaging, satellite monitoring, and mobile photography — captured images suffer from low resolution, noise, blur, and unknown degradation processes. Enhancing low-quality images into high-resolution, visually faithful outputs is a core challenge in computer vision and generative AI.

Core Objective

Participants must build a model that takes Low-Resolution (LR) images as input and outputs High-Resolution (HR) enhanced images, upscaling by a factor of ×2, handling unknown degradation operators.

Dataset Description

The dataset consists of paired images: High-Resolution Images (HR) as ground truth targets, and Low-Resolution Images (LR) generated using unknown degradation operators including Gaussian blur, noise injection, compression artifacts, and mixed distortions.

Participants must learn the degradation mapping implicitly.

Challenge Tracks

Track 2 – Unknown Degradation (×2)

Upscale LR images by a factor of 2 and reconstruct high-quality HR outputs. No knowledge of the degradation process will be provided.

Allowed Approaches

CNN-based Models

  • • SRCNN
  • • EDSR
  • • RCAN

GAN-based Methods

  • • SRGAN
  • • ESRGAN

Transformer-based Models

  • • SwinIR
  • • Vision Transformers

Diffusion-based Models

Pretrained models are allowed but must be disclosed.

Evaluation Metrics

Objective Metrics

  • • PSNR (Peak Signal-to-Noise Ratio)
  • • SSIM (Structural Similarity Index)
  • • Higher is better

Perceptual Quality (Bonus)

  • • LPIPS (Learned Perceptual Image Patch Similarity)
  • • Human evaluation score

Scoring Breakdown

PSNR
40%
SSIM
30%
Generalization
15%

Rules & Constraints

Key Rules

  • • Only provided training data may be used
  • • Test ground truth HR images will not be provided
  • • Inference must be reproducible
  • • Models must upscale exactly ×2
  • • No manual editing of outputs

Submission Requirements

  • • Enhanced HR images for the validation/test LR inputs
  • • Model architecture description
  • • Training methodology
  • • Hyperparameter configuration
  • • Runtime and hardware specifications
  • • Short technical report (max 8 pages)