The First Challenge on Mobile Real-World Image Super-Resolution at NTIRE 2026: Benchmark Results and Method Overview

1. Problem Statement

The paper addresses the critical gap between high-performance image super-resolution (SR) and mobile deployability.

• The Core Challenge: Reconstructing High-Resolution (HR) images from Low-Resolution (LR) inputs with a $\times4$ scaling factor under unknown, real-world degradations (e.g., sensor noise, compression artifacts, unknown blur).
• The Constraint: Unlike traditional SR benchmarks that assume simple bicubic downsampling, this challenge targets mobile devices. Models must be lightweight enough to run on resource-constrained hardware (specifically the MediaTek Dimensity 8400) with strict limits on parameters, FLOPs, and inference latency, while maintaining high perceptual quality.
• The Trade-off: Modern generative models (Diffusion, GANs) offer superior perceptual realism but are often too heavy for real-time mobile inference. The challenge seeks solutions that balance perceptual quality with computational efficiency.

2. Methodology & Evaluation Framework

Dataset

• Training Data: DIV2K (800 images) and LSDIR (84,991 images). Participants could use external data (e.g., Flickr2K).
• Degradation: LR-HR pairs were generated using a $\times4$ downsampling factor with complex, unknown degradation pipelines to simulate real-world conditions.
• Testing: 100 LR images from the DIV2K validation set (HR ground truth hidden) and 100 from LSDIR.

Evaluation Metrics

The final score is a weighted combination of Perceptual Quality and Inference Speed:

1. Perceptual Score: A composite metric based on six Image Quality Assessment (IQA) algorithms:
   • Full-reference: LPIPS, DISTS.
   • No-reference: CLIP-IQA, MANIQA, MUSIQ, NIQE.
   • Formula: $Score = (1-LPIPS) + (1-DISTS) + CLIP-IQA + MANIQA + \frac{MUSIQ}{100} + \max(0, \frac{10-NIQE}{10})$.
2. Speedup Ratio: Measured on the MediaTek Dimensity 8400 (FP16 precision).
   • $Speedup = \frac{t_{OSEDiff}}{t_{current\_model}}$, where $OSEDiff$ is the baseline.
3. Final Score: $FinalScore = 2 \times Score \times Speedup^{0.2}$.

3. Key Contributions & Trends

The paper analyzes 16 valid teams and identifies five dominant technical trends that emerged to solve the mobile SR problem:

1. One-Step Diffusion Dominance:

   • Teams moved away from multi-step diffusion (too slow) toward one-step diffusion frameworks (e.g., OSEDiff, TADSR, PiSA-SR).
   • These models leverage pre-trained Stable Diffusion priors but compress the generation process into a single inference step.

2. Knowledge Distillation & Compression:

   • To meet mobile constraints, teams employed distillation strategies, transferring knowledge from heavy "teacher" models to lightweight "student" networks.
   • Techniques included adversarial diffusion compression, aggressive UNet pruning, and lightweight decoders (e.g., TAESD).

3. Latent Space & Timestep Optimization:

   • To stabilize one-step diffusion, teams introduced Latent-Timestep Alignment (LSA) and Asynchronous Conditional Score Distillation (ACSD).
   • These methods anchor low-quality inputs to the diffusion manifold more accurately, reducing hallucinations and structural distortions.

4. Hybrid Architectures:

   • Teams combined the structural stability of GANs (e.g., Real-ESRGAN) with the texture richness of Diffusion models.
   • Ensemble methods (e.g., weighted blending of GAN and Diffusion outputs) were used to maximize perceptual metrics while maintaining robustness.

5. IQA-Aware Training Curricula:

   • Training shifted from pixel-level loss (L1/L2) to multi-stage fine-tuning guided by perceptual metrics.
   • Pipelines often involved: (1) Base adaptation on synthetic data, (2) Semantic LoRA fine-tuning, and (3) Score-oriented refinement using differentiable IQA losses (MANIQA, MUSIQ, etc.).

4. Results

• Participation: 108 registrants, with 16 teams achieving valid scores.
• Top Performers:
  • 1st Place (VIPSL): Used a compact PLKSR-Rep backbone with IQA-aware optimization and multi-stage fine-tuning.
  • 2nd Place (Antman): Utilized a simple RRDBNet (ESRGAN) but optimized the objective function with perceptual score guidance.
  • 3rd Place (SamsungAICamera): Distilled a one-step diffusion teacher (OSEDiff) into a pruned student with a shallow Detail Enhancement Module (DEM).
• Performance: The top teams achieved a significant trade-off, with the winner (VIPSL) achieving a perceptual score of 4.03 and a speedup of 113.85x over the baseline, resulting in a final score of 42.21.

5. Significance

• Bridging the Gap: This challenge successfully demonstrated that state-of-the-art perceptual quality is achievable on mobile hardware without relying on massive, impractical models.
• New Paradigm: It established one-step diffusion combined with knowledge distillation as the new standard for mobile real-world SR, moving beyond traditional CNNs and GANs.
• Practical Deployment: By enforcing strict hardware constraints (MediaTek Dimensity 8400) and using real-world degradation, the results provide a realistic benchmark for deploying AI image enhancement in consumer smartphones and edge devices.
• Community Resource: The paper provides a comprehensive overview of methods, code, and pre-trained models, fostering further research in efficient generative vision.

In summary, the NTIRE 2026 Mobile Real-World SR challenge proved that through architectural compression, distillation, and perceptual-aware training, high-fidelity image restoration can be executed efficiently on mobile devices.