Semantically-Aware Game Image Quality Assessment

TL;DR

This work targets no-reference image quality assessment for video game graphics, addressing the unique distortions and lack of reference frames that standard NR-IQA/VQA methods struggle with. The authors propose a semantically-aware pipeline that combines a knowledge-distilled Game Distortion Feature Extractor (GDFE) with CLIP-based semantic gating to adapt distortion importance to scene content, trained on game presets and distortion datasets from Forza Horizon 5. Key contributions include a two-dataset training regime (Distortion and Preset), a binary-teacher knowledge-distillation scheme that yields a robust GDFE for intermediate distortions, and a semantically gated quality regressor that achieves stable, monotonic predictions across unseen titles in the same genre. The approach demonstrates strong generalization within racing games, reduces output variance, and provides a foundation for automated graphical quality assessment and adaptive rendering configurations in gaming.

Abstract

Assessing the visual quality of video game graphics presents unique challenges due to the absence of reference images and the distinct types of distortions, such as aliasing, texture blur, and geometry level of detail (LOD) issues, which differ from those in natural images or user-generated content. Existing no-reference image and video quality assessment (NR-IQA/VQA) methods fail to generalize to gaming environments as they are primarily designed for distortions like compression artifacts. This study introduces a semantically-aware NR-IQA model tailored to gaming. The model employs a knowledge-distilled Game distortion feature extractor (GDFE) to detect and quantify game-specific distortions, while integrating semantic gating via CLIP embeddings to dynamically weight feature importance based on scene content. Training on gameplay data recorded across graphical quality presets enables the model to produce quality scores that align with human perception. Our results demonstrate that the GDFE, trained through knowledge distillation from binary classifiers, generalizes effectively to intermediate distortion levels unseen during training. Semantic gating further improves contextual relevance and reduces prediction variance. In the absence of in-domain NR-IQA baselines, our model outperforms out-of-domain methods and exhibits robust, monotonic quality trends across unseen games in the same genre. This work establishes a foundation for automated graphical quality assessment in gaming, advancing NR-IQA methods in this domain.

Figures (12)

• Figure 1: An section of the extensive graphical settings menu in Cyberpunk 2077, illustrating the complexities of configuration options across multiple pages.
• Figure 2: Predicted scores from each of COVER’s semantic, technical, and aesthetic branches, as well as the aggregated overall score, for a video dataset collected at low, medium, high, and ultra qualities in Forza Horizon 5. The predicted scores show no correlation with the quality presets (SRCC < 0.2). Ideally, the predicted scores should increase monotonically from low to ultra qualities. The lack of correlation indicates that COVER is ineffective in the gaming domain.
• Figure 3: Predicted distortion scores for Forza Horizon 5 validation video data for UVQ’s 25 distortion types, along with an additional class for non-distortion. No monotonic trend is observable across the quality levels for each of the distortions, corroborated by the low correlation scores. JPEG compression evaluations output 0 across all quality levels, since the videos are evaluated at raw capture quality with no compression.
• Figure 4: Comparison of the directly-trained GDFE classifier (Model a) and the knowledge-distilled GDFE regressor (Model b) performance on FH5 test data. In this comparison, higher distortion severity values correspond to lower visual quality. Model (a) fails to predict a significant correlation between the input features and visual quality, whereas Model (b) successfully predicts an inverse correlation between distortion severity and quality. Section 5.1 provides a detailed discussion of these trends.
• Figure 5: An overview of the game quality assessment model, featuring the game distortion feature extractor (GDFE) trained via knowledge distillation using soft labels from teacher binary classifiers. The GDFE’s feature maps are gated using image-text similarity to produce semantic-aware quality scores.