Automatic Quality Assessment of AIGC 3D Game Assets for Production Readiness and Its In-Engine Implementation

Authors

  • Danyang Wang School of Digital Media, Guangzhou University of Software, Guangzhou, Guangdong, 510990, China
  • Xianyuan Liu College of Electronic Information and Control Engineering, Guangzhou University of Software, Guangzhou, Guangdong, 510990, China

DOI:

https://doi.org/10.54691/5ncsz938

Keywords:

AIGC; 3D Game Asset; Quality Assessment; Defect Injection; Geometric Descriptors; Unreal Engine.

Abstract

AIGC-based 3D generators have lowered the barrier of game asset production, yet their outputs frequently contain production-blocking defects such as non-manifold edges, flipped normals and overlapping UVs, while studios still rely on manual visual inspection. This paper presents a production-oriented quality assessment method for 3D game assets together with an in-engine implementation. A balanced 120-sample validation set with four severity levels is constructed by gradient defect injection on six kinds of procedural primitives. Twenty-nine geometric descriptors are extracted to train Random Forest and SVM-RBF classifiers, which are compared with a pure rule-based baseline. A dual-channel architecture that combines interpretable geometric rules with multi-view semantic scoring is designed, and the semantic channel is calibrated with real CLIP ViT-B/32 weights on CPU. The whole pipeline is encapsulated as a UE5 Python editor plugin. To prevent variants of the same prototype from crossing folds, repeated five-fold cross-validation grouped by prototype is adopted. Random Forest reaches 88.00%±0.70% accuracy and 0.8792±0.0071 F1 on the usable/unusable task, against 66.67% for the rule baseline; the grouped protocol is 2.67 percentage points higher than the stratified one, so no leakage-driven inflation is observed, and McNemar's test finds no significant difference from SVM-RBF (85.83%, p=0.581). The proportion of samples judged usable decreases monotonically across the four severity levels (100%, 73.3%, 20.0% and 0%). Winding inconsistency ratio (0.2452), inward-face ratio (0.1273) and open-edge ratio (0.1271) are the most informative descriptors. The semantic channel shows two-sided behaviour: text-only zero-shot scoring is not discriminative (AUC 0.27, with scores rising as defects worsen), whereas the multi-view embedding distance to clean reference renders reaches AUC 0.899 and 84.2% accuracy, approaching the geometric channel (AUC 0.919). Fusing both channels raises AUC only to 0.9228, indicating that they are largely redundant rather than complementary. CLIP encoding costs 73.31 ms per image (224x224, CPU, batch 8). In a real UE5 5.4.4 project, the plugin enumerates 1873 assets, loads 223 static meshes, and completes the export-inference-metadata-writeback-verification loop for 8 samples with an 8/8 success rate in 6.52 s. All reported separability is confined to the synthetic injection protocol and must be re-examined on real generated assets.

Downloads

Download data is not yet available.

References

[1] Poole, B., Jain, A., Barron, J. T., & Mildenhall, B. (2023). DreamFusion: Text-to-3D using 2D diffusion. In Proceedings of the International Conference on Learning Representations.

[2] Jun, H., & Nichol, A. (2023). Shap-E: Generating conditional 3D implicit functions. arXiv:2305.02463.

[3] Fukaya, K., Daylamani-Zad, D., & Agius, H. (2024). Evaluation metrics for intelligent generation of graphical game assets: A systematic survey-based framework. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12), 7998–8017. https://doi.org/10.1109/TPAMI.2024.3400447.

[4] Fu, K., Duan, H., Zhang, Z., et al. (2025). Multi-dimensional quality assessment for text-to-3D assets: Dataset and model. IEEE Transactions on Multimedia, 27, 8145–8159. https://doi.org/10.1109/TMM.2025.3478612.

[5] Duggal, S., Hu, Y., Michel, O., et al. (2025). Eval3D: Interpretable and fine-grained evaluation for 3D generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 13326–13336).

[6] Asaria, A., Salomone, T., & Gandhi, D. (2026). A cross-model VLM-judge protocol for single-image 3D mesh quality. arXiv:2606.18451.

[7] Radford, A., Kim, J. W., Hallacy, C., et al. (2021). Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning (pp. 8748–8763).

[8] Liu, H., Li, C., Wu, Q., & Lee, Y. J. (2023). Visual instruction tuning. In Advances in Neural Information Processing Systems.

[9] Bai, J., Bai, S., Yang, S., et al. (2023). Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv:2308.14466.

[10] Dawson-Haggerty, E., et al. (2019). trimesh: Python library for triangular meshes. https://github.com/mikedh/trimesh.

[11] Pedregosa, F., Varoquaux, G., Gramfort, A., et al. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825–2830. https://doi.org/10.5555/1953048.2078195.

[12] Hu, E. J., Shen, Y., Wallis, P., et al. (2022). LoRA: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations.

Downloads

Published

2026-11-20

Issue

Section

Articles

How to Cite

Wang, D., & Liu, X. (2026). Automatic Quality Assessment of AIGC 3D Game Assets for Production Readiness and Its In-Engine Implementation. Scientific Journal of Technology, 8(11), 84-94. https://doi.org/10.54691/5ncsz938