Overview

Recent advances in Artificial Intelligence Generated Content (AIGC) have transformed synthetic image creation. Diffusion- and transformer-based generative models such as DALL·E, Stable Diffusion, and Imagen can now generate visually compelling, semantically rich, and diverse images across a wide range of applications. Despite this progress, AIGC images often contain non-traditional distortions, including semantic inconsistencies, unnatural object structures, and perceptual artifacts, that are poorly handled by conventional Image Quality Assessment (IQA) methods.

Most existing IQA models are designed for traditional degradations such as blur, noise, or compression artifacts and rely heavily on large pretrained backbones and external data. This makes them unsuitable for evaluating AIGC content and impractical for resource-constrained and edge deployment scenarios.

The ICME 2026 Grand Challenge on Lightweight Neural Networks for AIGC Image Quality Assessment addresses these limitations by encouraging the development of compact, no-reference IQA models that are trained strictly from scratch and optimized for extreme efficiency while maintaining high correlation with human perceptual judgments.

🎯 Challenge Objectives

Participants are expected to:

• Design lightweight neural network architectures for no-reference AIGC IQA

• Train models from scratch without using pretrained weights or external data

• Achieve strong perceptual correlation under strict parameter and efficiency constraints

• Explore perceptually-informed feature design, efficient learning strategies, and optimization techniques

The challenge promotes fair benchmarking, deployable solutions, and innovation aligned with ICME’s focus on efficient multimedia processing and edge intelligence.

📊 Datasets

The challenge leverages a diverse set of publicly available AGIQA datasets, covering multiple styles, scales, and content types:

• AGIQA-1K: 1,080 AI-generated images with Mean Opinion Scores (MOS)

• AGIQA-3K: ~3,000 images with fine-grained subjective annotations

Standardized dataset splits will be provided:

• Training set (labeled)

• Private test set (used for final ranking)

Note: Participants are encouraged to use AGIQA-1K and AGIQA-3K freely for training and internal validation as they prefer.

🧪 Task Definition

• Task: No-reference perceptual quality prediction for AIGC images

• Input: Single AI-generated image

• Output: Predicted perceptual quality score

⚙️ Model Constraints & Rules

• ❌ No pretrained models

• ❌ No external datasets

• ✅ Training strictly from scratch

• 📦 Model size ≤ 10 million parameters

• 👥 Teams of up to 5 members (1 person can not be part of multiple teams).

• 🔁 Full reproducibility required (code + model + executable submission)

📈 Evaluation Criteria

Submissions are evaluated based on prediction accuracy, model efficiency, and technical novelty, emphasizing lightweight, reproducible models.

Primary Metrics (Private Test Set)

• Pearson Linear Correlation Coefficient (PLCC)

• Spearman Rank-Order Correlation Coefficient (SRCC)

• Kendall Rank Correlation Coefficient (KRCC)

• Root Mean Square Error (RMSE)

Model Performance Score

Correlation Score = (max(PLCC,0) + max(SRCC,0) + max(KRCC,0)​) / 3

Model Efficiency (Size) Score

The model size contribution evaluates the efficiency of submitted models using a tiered, percentile-based scoring system:

Model size percentile  - Score contribution (out of 20)

Top 10% (smallest) - 20

Top 25% - 18

Top 50% - 15

Top 75% - 10

Bottom 25% - 5

• Smaller models receive higher scores, incentivizing lightweight, deployable designs.

• Scores are assigned relative to all submissions, making the system fair and adaptive.

• Edge cases near percentile boundaries are smoothed linearly to avoid sudden jumps in points, ensuring small differences in size do not disproportionately affect the score.

• This approach encourages participants to optimize both accuracy and efficiency, balancing model performance with compactness.

Paper Innovation Score

Short papers (max 4 pages) are evaluated based on:

• Innovation & originality

• Technical rigor and methodology

• Alignment with challenge objectives

Final Ranking Weights

• Test set performance using the submitted model (code): 50%

• Model efficiency (size and compactness): 20%

• Novelty and technical quality of the proposed method (4-page short paper): 30%

Ties are broken using RMSE and model size.

Participants are encouraged to explore techniques such as pruning, quantization, and other efficient neural network design methods while maintaining strong alignment with human perceptual judgments. Additional techniques are permitted only if the overall setup and model comply with all the above requirements.

🏆 Awards & Recognition (ICME26 AIGC-IQA-GC)

🥇 Winner Award

• Awarded to the team achieving the highest overall ranking score

• Based on the combined evaluation of performance, efficiency, and innovation

🥈 Runner-up Award

• Awarded to the team with the second-highest overall ranking score

⚡ Most Efficient Design Award

• Awarded to the team with the best model efficiency (smallest models)

• Eligible submissions must achieve at least top 50% performance (Composite Score)

• Highlights lightweight, deployable solutions without sacrificing accuracy

📊 Best Performance Design Award

• Awarded to the team with the highest test performance (Composite Score)

• Eligible submissions must fall within the top 50% smallest models

• Encourages high accuracy under reasonable efficiency constraints

💡 Most Innovative Design Award

• Awarded based on novelty, originality, and technical quality of the proposed method (short paper)

• Recognizes creative approaches and impactful ideas aligned with the challenge goals

Important Dates (tentative)

• Mar 15, 2026 Challenge registration opens [Register Here]

• Mar 22, 2026 Training data, baseline code & evaluation scripts released

• Apr 21, 2026 Challenge closes

• Apr 23, 2026 Challenge Paper submission deadline

• Apr 30, 2026 Paper acceptance notification

• May 10, 2026 Camera-ready submission

• Jul 5-9, 2026 ICME 2026 Grand Challenge session & winner announcement

Organizing Committee

🤝 Sponsorship & Support

Industry partners are welcome to sponsor prizes, awards, or travel grants. Please contact the organizers for collaboration opportunities.

Contact Email

If you have any questions or concerns, please contact us: icme26-aigciqa-gc@googlegroups.com 

Join us at ICME 2026 and contribute to shaping the future of efficient and perceptually-aware AIGC quality assessment!