Automated Detection of 3D Game Visual Defects based on a Lightweight Vision-Language Model
DOI:
https://doi.org/10.54691/7mdcxj65Keywords:
Vision-language Model; Game Visual Defect Detection; LoRA Fine-tuning; Prompt Engineering; Local Deployment.Abstract
Modern 3D games use complex rendering pipelines that make manual visual quality assurance expensive and difficult to scale. This study presents an automated screenshot-based defect detection method using a lightweight vision-language model. Qwen2.5-VL-7B is adapted with LoRA on a small domain dataset covering lighting, texture/material, and post-processing anomalies. Prompt constraints and Operational Design Domain cleaning are introduced to distinguish genuine defects from acceptable effects such as bloom, fog, and strong daylight. The adapted model is merged, converted to GGUF, quantized with Q4_K_M, and deployed offline through llama.cpp on an RTX 4060. On a 176-image evaluation set, the finalized configuration achieves 80.68% accuracy, 81.25% precision, and 70.27% recall, while reducing false positives to 12. The results indicate that few-shot adaptation and explicit task-boundary design can provide a low-cost, privacy-preserving pre-screening tool for 3D game quality assurance.
Downloads
References
[1] Politowski, C., Petrillo, F., & Guéhéneuc, Y. G. (2021). A survey of video game testing. arXiv preprint arXiv:2103.06431.
[2] Păduraru, C. I., Păduraru, M. G., & Stefanescu, A. (2025). Hierarchical deep learning framework for continuous, state-aware visual glitch detection in games. In Evaluation and Assessment in Software Engineering (EASE 2025) (Istanbul, Türkiye). https://doi.org/10.1145/3756681.3757010
[3] Paduraru, C. (2026). A state-aware, hierarchical deep learning framework for automated visual glitch detection in games. Engineering Applications of Artificial Intelligence, 166, 113497. https://doi.org/10.1016/j.engappai.2025.113497
[4] Zhou, T. (2025). GlitchAgent: Detecting Video Game Glitches from Gameplay Videos [Master’s thesis]. Virginia Polytechnic Institute and State University, Blacksburg.
[5] Liu, H., Li, C., Wu, Q., & Lee, Y. J. (2023). Visual instruction tuning. In Advances in Neural Information Processing Systems (Vol. 36).
[6] Chu, X., Qiao, L., Lin, X., et al. (2023). MobileVLM: A fast, strong and open vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886.
[7] Hu, S., Tu, Y., Han, X., et al. (2024). MiniCPM: Unveiling the potential of small language models with scalable training strategies. arXiv preprint arXiv:2404.06395.
[8] Jeong, J., Zou, Y., Kim, T., et al. (2023). WinCLIP: Zero-/few-shot anomaly classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 19606–19616).
[9] Liu, Z., Chen, C., Wang, J., et al. (2020). Owl eyes: Spotting UI display issues via visual understanding. In 35th IEEE/ACM International Conference on Automated Software Engineering (Virtual Event, Australia). https://doi.org/10.1145/3324884.3416547
[10] Hu, E. J., Shen, Y., Wallis, P., et al. (2022). LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.
[11] Qwen Team. (2025). Qwen2.5-VL technical report. arXiv preprint arXiv:2502.13923.
Downloads
Published
Issue
Section
License
Copyright (c) 2026 Scientific Journal of Intelligent Systems Research

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.




