This literature-based review examines recent developments in deep learning–based hybrid approaches for enhancing photorealism in real-time game engine rendering. Rather than replacing conventional rendering pipelines, the reviewed studies converge on augmenting them with learned post-processing stages. Accordingly, the scope of this review is limited to methods that operate on rendered frames and engine-generated auxiliary representations, rather than fully neural scene synthesis. The central research problem addressed across the literature is the persistent “appearance gap” between real-world imagery and real-time rendering, which arises from the trade-off between computational efficiency and visual fidelity.
Traditional rendering pipelines rely on physically based simulation techniques, such as rasterization and path tracing, to approximate light transport and material behavior. While path tracing achieves high physical accuracy, it remains computationally prohibitive for real-time applications. Conversely, rasterization-based approaches are efficient but often produce visually incomplete results. This trade-off has led to the emergence of hybrid methods, which preserve the geometric, material, and lighting information already computed by game engines while applying data-driven enhancement techniques to improve visual realism. In this context, photorealism is reframed as a post-render optimization problem rather than a purely physics-based simulation challenge.
A key methodological trend identified in the literature is the use of G-buffers as structured inputs for deep learning models. These buffers encode scene-level information such as depth, surface normals, albedo, and material properties, enabling models to operate with physically meaningful context rather than relying solely on RGB inputs or semantic labels. This approach improves stability and reduces ambiguity in image enhancement tasks, as it allows networks to preserve geometric and material coherence during reconstruction. Convolutional neural networks (CNNs), often combined with encoder–decoder architectures and multi-scale feature extraction, are widely used in this context. Additionally, rendering-aware modules such as rendering-aware denormalization (RAD) have been introduced to integrate scene information directly into feature modulation processes.
Adversarial training also remains a central component of hybrid photorealism techniques. However, it is typically employed as part of a broader perceptual learning framework, often combined with semantic guidance and patch-based sampling strategies. These methods aim to reduce artifacts and improve visual plausibility, particularly in complex urban environments. Closely related to these approaches is neural supersampling, where low-resolution frames are enhanced using temporal information, motion vectors, and previous frames. This allows high-resolution outputs to be generated without incurring the full computational cost of native rendering.
The literature further highlights the importance of dataset characteristics in determining model performance. Real-world datasets such as KITTI, Cityscapes, and Mapillary Vistas differ not only stylistically but also structurally in terms of scene layout, sensor properties, and environmental features. These differences significantly influence model generalization and failure modes. In particular, mismatches between synthetic game data and real-world distributions can lead to visual inconsistencies and hallucinations, such as semantically implausible object placements. As a result, buffer-aware and semantically guided approaches consistently outperform naïve image-to-image translation methods.
Despite these challenges, the reviewed studies demonstrate that hybrid approaches provide meaningful improvements in both visual quality and computational efficiency. Reported results include enhanced reflections on vehicle surfaces, improved vegetation reconstruction aligned with target dataset characteristics, and more realistic road textures. These improvements indicate that deep learning models can reinterpret engine outputs in ways that better approximate real-world visual appearance. Furthermore, recent research emphasizes deployable system design, proposing two-stage pipelines in which high-quality offline models are used to generate training data for lightweight real-time models. Such approaches enable significant performance gains while maintaining acceptable visual fidelity.
Nevertheless, hybrid photorealism enhancement remains an open research problem. Temporal inconsistencies, particularly in high-motion scenes, continue to challenge sequence-level coherence. Additionally, material representation errors such as exaggerated reflections or texture distortions can reduce perceptual realism. Computational constraints also persist, as high-resolution enhancement and adversarial models still impose nontrivial latency and memory requirements. Another limitation concerns evaluation: the lack of standardized metrics across studies complicates direct comparison and hinders the establishment of production-level benchmarks.
In conclusion, the literature suggests that hybrid deep learning approaches are evolving into a critical component of modern rendering pipelines. Their success depends not only on visual quality but also on temporal stability, computational efficiency, and integration within real-time systems. Future research directions include improving temporal consistency, enhancing cross-dataset generalization, developing lightweight architectures that retain buffer awareness, and establishing unified evaluation frameworks that jointly consider realism, performance, and stability.
Bu çalışma, bilgisayar oyunlarında gerçek zamanlı render çıktılarının fotogerçekçi hale getirilmesine yönelik derin öğrenme tabanlı hibrit yaklaşımları literatür temelli bir derleme kapsamında incelemektedir. Geleneksel grafik işleme hatlarının (rendering pipelines) fiziksel doğruluk ile hesaplama maliyeti arasında kurmak zorunda kaldığı denge, özellikle gerçek zamanlı uygulamalarda görsel kaliteyi sınırlayan önemli bir faktör olarak öne çıkmaktadır. Bu durum, literatürde sıklıkla “görünüm farkı” (appearance gap) olarak tanımlanmakta ve son yıllarda bu farkın azaltılmasına yönelik veri odaklı yöntemlere olan ilgiyi artırmaktadır.
Bu bağlamda, evrişimsel sinir ağları (CNN) ve karşıtlık eğitimi (adversarial training) gibi derin öğrenme tekniklerinin, oyun motoru çıktılarıyla birlikte kullanıldığı hibrit yaklaşımlar ele alınmaktadır. Özellikle, görüntülerin yalnızca anlamsal etiket haritalarından yeniden sentezlenmesi yerine; oyun motorları tarafından üretilen geometri, yüzey normalleri, materyal özellikleri ve ışıklandırma gibi yardımcı verilerin derin öğrenme modellerine girdi olarak sunulması, sahne yapısının korunmasını sağlayarak görsel kaliteyi artıran önemli bir yöntem olarak değerlendirilmektedir.
Çalışma kapsamında, Unreal Engine ve GTA V gibi ortamlardan elde edilen sentetik görüntülerin, KITTI, Cityscapes ve Mapillary Vistas gibi gerçek dünya veri setleriyle eğitilmiş modeller aracılığıyla nasıl iyileştirildiğine ilişkin literatürde yer alan çalışmalar karşılaştırmalı olarak incelenmektedir. Bu çalışmaların bulguları, derin öğrenme tabanlı yöntemlerin araç yüzeylerindeki yansımaları geliştirme, bitki örtüsünü veri setine özgü özelliklere göre yeniden yapılandırma ve yol yüzeylerinde daha gerçekçi dokular üretme gibi katkılar sağladığını göstermektedir.
Bununla birlikte, zamansal tutarsızlıklar, veri setleri arasındaki yapısal farklılıklar ve model genelleme problemleri gibi teknik sınırlılıklar da literatürde vurgulanmaktadır. Genel olarak değerlendirildiğinde, veri odaklı hibrit yaklaşımların geleneksel render süreçlerini tamamlayıcı bir rol üstlendiği ve fotogerçekçilik üretiminde önemli bir potansiyel sunduğu görülmektedir. Bu çalışma, söz konusu yöntemleri kavramsal ve uygulama temelli literatür üzerinden sistematik biçimde analiz ederek alandaki güncel eğilimlere bütüncül bir bakış sunmayı amaçlamaktadır.
Akenine-Möller, T., Haines, E., Hoffman, N., Pesce, A., Iwanicki, M., ve Hillaire, S. (2018). Real-time rendering (4. bs.). A K Peters/CRC Press.
Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., ve Schiele, B. (2016). The Cityscapes dataset for semantic urban scene understanding. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 3213–3223.
Geiger, A., Lenz, P., Stiller, C., ve Urtasun, R. (2013). Vision meets robotics: The KITTI dataset. International Journal of Robotics Research, 32(11), 1231–1237. https://doi.org/10.1177/0278364913491297
Jetchev, N., Bergmann, U., ve Vollgraf, R. (2016). Texture synthesis with spatial generative adversarial networks. arXiv. https://arxiv.org/abs/1611.08207
Mameli, M., Paolanti, M., Mancini, A., Zingaretti, P., ve Pierdicca, R. (2025). RenderGAN: Enhancing real-time rendering efficiency with deep learning. ACM Transactions on Multimedia Computing, Communications, and Applications, 21(3), Makale 99. https://doi.org/10.1145/3712263
Neuhold, G., Ollmann, T., Bulò, S. R., ve Kontschieder, P. (2017). The Mapillary Vistas dataset for semantic understanding of street scenes. International Conference on Computer Vision (ICCV).
Newell, M. E., ve Blinn, J. F. (1977). The progression of realism in computer generated images. Proceedings of the ACM Annual Conference, 446–451. https://doi.org/10.1145/800179.810236
Pasios, S., ve Nikolaidis, N. (2025). REGEN: Real-time photorealism enhancement in games via a dual-stage generative network framework. arXiv. https://arxiv.org/abs/2508.17061
Richter, S. R., Abu AlHaija, H., ve Koltun, V. (2023). Enhancing photorealism enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2), 1700–1715. https://doi.org/10.1109/TPAMI.2021.3082179
Sancaklı, P., ve Sarı, Ü. (2025). Dijital oyunlarda yapay zekâ kullanımı ve ‘F.E.A.R’ oyunu örneğinde güncel yönelimler. Communicata, 29, 87–98. https://doi.org/10.32952/communicata.1628205
Stacchio, L., Balloni, E., Gorgoglione, L., Mancini, A., Giovanola, B., Tiribelli, S., ve Zingaretti, P. (2024). An ethical framework for trustworthy neural rendering applied in cultural heritage and creative industries. Frontiers in Computer Science, 6, 1459807. https://doi.org/10.3389/fcomp.2024.1459807
Weyrich, T., Lawrence, J., Lensch, H. P. A., Rusinkiewicz, S., ve Zickler, T. (2009). Principles of appearance acquisition and representation. Foundations and Trends in Computer Graphics and Vision, 4(2), 75–191. https://doi.org/10.1561/0600000022
Wu, Y., Yi, A., Ma, C., ve Chen, L. (2023). Artificial intelligence for video game visualization, advancements, benefits and challenges. Mathematical Biosciences and Engineering, 20(8), 15345–15373. https://doi.org/10.3934/mbe.2023686
Xiao, L., Nouri, S., Chapman, M., Fix, A., Lanman, D., ve Kaplanyan, A. (2020). Neural supersampling for real-time rendering. ACM Transactions on Graphics, 39(4), Makale 142. https://doi.org/10.1145/3386569.3392376
