Intel Unveils AI-Powered Tool for Real-Time Gaming Image Quality Assessment

Intel's Innovative Approach to Gaming Image Quality Assessment

Intel has unveiled a groundbreaking AI-powered tool designed to objectively evaluate image quality in modern video games. The Computer Graphics Visual Quality Metric (CGVQM), now available on GitHub as a PyTorch application, aims to address the challenges posed by contemporary rendering techniques 1

1

3

The Need for Advanced Image Quality Metrics

Modern games rarely render frames natively, instead relying on techniques such as upscalers (like DLSS) and frame generation. These methods can introduce various image quality issues, including ghosting, flickering, and aliasing. While these problems are often discussed qualitatively, assigning objective measurements has been challenging 1

1

Existing metrics like peak signal-to-noise ratio (PSNR) have limitations when applied to real-time graphics output. To overcome these constraints, Intel researchers developed a two-pronged approach:

1. Creation of a new video dataset (CGVQD) showcasing various image quality degradations.
2. Training of an AI model (CGVQM) to rate image quality based on these distortions 1

   1

   [2].

The CGVQM Model: A Technological Breakthrough

The CGVQM model utilizes a 3D convolutional neural network (CNN) architecture, specifically a 3D-ResNet-18 model. This choice allows the network to consider both spatial and temporal pattern information, crucial for high-performance image quality evaluation 1

1

Key features of the CGVQM model include:

• Outperforms other image quality evaluation tools on the researchers' dataset
• Generalizes well to videos outside its training set
• Provides per-pixel error maps and attempts to identify the cause of artifacts 1

  1

  [2]3

  3

Source: Tom's Hardware

Related Stories

Nvidia's DLSS 5 AI tool sparks fierce backlash over beauty filter effects on game characters

Intel brings XeSS 3.0 frame generation to older Arc GPUs, but keeps code proprietary

Nvidia unveils three AI models for DLSS 5, promises developers control over artistic intent

NVIDIA's DLSS Adoption Soars: Over 80% of GeForce RTX Users Embrace AI Upscaling

Human-Aligned Performance and Real-World Applications

To ensure the AI model's observations aligned with human perception, researchers conducted a study with 20 participants. These human observers rated various distortions in video sequences, establishing a baseline for the AI model 1

1

3

The CGVQM tool's ability to predict human judgment of visual distortions makes it valuable for:

• Optimizing quality and performance trade-offs in upscalers
• Providing smarter reference generation for training denoising algorithms
• Evaluating engine updates and new upscaling techniques 3

  3

Future Developments and Limitations

While the current version of CGVQM shows promise, there's room for improvement:

• Potential use of transformer neural network architecture for enhanced performance
• Incorporation of optical flow vectors to refine image quality evaluation
• Ongoing work to include saliency, motion coherence, and semantic awareness 1

  1

  [2]3

  3

Source: PC Gamer

The tool's reliance on reference videos currently limits some applications. However, Intel's researchers are working to expand CGVQM's capabilities, making it more robust for real-world scenarios 3

3