Intel has introduced the Computer Graphics Visual Quality Metric (CGVQM), an open-source AI-powered tool available on GitHub, designed to objectively evaluate image quality in modern video games with high accuracy.

The dataset includes 80 three-second clips from open-source 3D scenes, providing a comprehensive basis for training the AI to detect issues such as ghosting, flicker, and aliasing caused by upscalers and frame generation.

This tool uses a 3D convolutional neural network architecture to analyze both spatial and temporal patterns, significantly improving its performance in assessing image quality.

Human observers rated the dataset to establish a baseline, ensuring the AI's evaluations align with human perception, with the CGVQM-5 model outperforming existing metrics and closely matching human assessments.

The CGVQM was trained on the CG-VQD dataset, which contains various degradations from rendering techniques like neural supersampling and path tracing, and was fine-tuned using a pre-trained 3D ResNet-18 network to closely match human quality ratings.

A lighter version, CGVQM-2, offers faster performance by using only part of the neural network, making it suitable for real-time applications.

Researchers see potential for further improvements, including incorporating transformer neural networks and optical flow information into the evaluation process.

The tool provides error maps that identify visual artifacts like ghosting and flicker, helping developers address quality issues efficiently during development.

Game developers can integrate CGVQM into their workflows via GitHub, adding Vulkan hooks or Unreal Engine plugins for real-time quality assessment.

Overall, the CGVQM marks a significant advancement in real-time graphics image quality evaluation, surpassing traditional metrics like PSNR, which are limited in this context.

The tool's performance is competitive with human judgment, making it a valuable asset for improving visual fidelity in modern gaming and rendering workflows.