NTIRE 2026 Mobile Real-World SR Challenge

NTIRE Workshop & Challenges @ CVPR 2026 • Denver, Colorado

Mobile Real-World Image Super-Resolution (×4)

Restore a high-resolution image from a single low-resolution input under unknown real-world degradations — and make it efficient enough for a mobile chipset.

Task
Mobile Real-World SR
Scale
×4
Focus
Perception + Mobile Latency

Challenge Overview

This track is part of the NTIRE 2026 Workshop & Challenges co-located with CVPR 2026 in Denver. It targets mobile real-world single-image super-resolution, emphasizing both visual quality and deployment efficiency.

What is NTIRE?

NTIRE (New Trends in Image Restoration and Enhancement) is a long-running workshop and challenge series promoting real-world datasets, standardized evaluation, and reproducible baselines across academia and industry.

Why mobile real-world SR?

Super-resolution must be more than “sharper” — it must be efficient, memory-friendly, and deployable. This track highlights both perceptual quality and on-device speed (measured on MediaTek Dimensity 8400).

Data & Competition Phases

Where to get the data

All datasets are hosted on the competition platform under the Files section.

Development: Training

Training pairs are provided. Teams train a ×4 SR model under unknown real-world degradations and prepare an efficient inference pipeline.

Development: Validation

Validation inputs are released (inputs only) to help you iterate and compare approaches during the development stage.

Test Phase

Final test inputs are released (inputs only). Participants submit SR results plus reproducible code/executables for official evaluation.

Evaluation

Perception Quality

A composite perceptual score is computed using a weighted combination of six no-reference (NR) quality metrics:

  • LPIPS
  • DISTS
  • NIQE
  • ManIQA
  • MUSIQ
  • CLIP-IQA

Mobile Efficiency

Inference speed is measured under FP16 on a real mobile platform:

  • MediaTek Dimensity 8400
  • Latency reported as a “Speedup” factor
  • Focus: practical deployability and real-time potential

Official Evaluation Setup

Organizers run your code/weights/inference script on:

  • 1× NVIDIA A6000 GPU to compute perceptual metrics and the overall score
  • MediaTek Dimensity 8400 for on-device speed

Determinism matters. Fix seeds and document every dependency.

Important Dates

2026-01-29
Training + Validation released
2026-03-10
Final test inputs released
2026-03-17
SR results + Fact Sheet + code/executable due
2026-03-19
Preliminary results announced
2026-03-24
Workshop paper deadline for challenge entries
2026-06
NTIRE Workshop & Awards (CVPR 2026, Denver)

Join the Challenge

Competition portal

Register, download data, and submit results on the CodaBench competition page.

Open CodaBench

NTIRE 2026

Learn more about the NTIRE 2026 workshop and other challenges on the official NTIRE website.

Open NTIRE 2026

Organizers

Jiatong Li
Shanghai Jiao Tong University
Jingkai Wang
Shanghai Jiao Tong University
Jue Gong
Shanghai Jiao Tong University
Zheng Chen
Shanghai Jiao Tong University
Kai Liu
Shanghai Jiao Tong University
Radu Timofte
University of Würzburg
Yulun Zhang
Shanghai Jiao Tong University

Organizer list may be updated.

WeChat Group

Scan the QR code below to join the official WeChat group for announcements and Q&A.

If the QR code expires, please contact the organizers by email.

WeChat group QR code
WeChat Group QR