From Pixels to Code: A Process for Generating Scalable Vector Game Art Assets Using Large Language Models

• Italo Thiago Felix Dos Santos UEA
• Luis Cuevas Rodriguez UEA
• Cristina Souza de Araujo UEA
• Jucimar Maia da Silva Junior UEA

DOI: https://doi.org/10.5753/sbgames.2026.26069

Resumo

The increasing demand for high-quality visual assets in indie and AAA games poses significant challenges for traditional raster-based generation methods, particularly due to limitations in scalability and resolution independence. Objective: This paper investigates the use of Large Language Models (LLMs) to support the generation of Scalable Vector Graphics (SVG) as game art assets, aiming to enhance production workflows. Methodology: We propose a process for code-based asset generation using locally deployed LLMs, including Qwen, Gemma, and GLM, constrained by rendering and structural rules to ensure valid SVG output. The generated assets are evaluated through quantitative metrics and statistical analysis, combined with a qualitative user study assessing aesthetic coherence, prompt fidelity, and usability for integration into game projects. The performance of local models is benchmarked against Gemini 3 as a strong proprietary baseline. Results: The results demonstrate that locally deployed models can generate structurally valid and scalable assets, with statistical evidence confirming consistent performance differences across models. However, a gap remains in aesthetic quality compared to the proprietary baseline. Conclusion: The findings indicate that LLM-based SVG generation is a viable and scalable approach for supporting iterative game art production workflows, highlighting its potential and limitations in aligning semantic intent with visual design quality.

Palavras-chave: Large Language Models, Scalable Vector Graphics (SVG), Game Assets, Local Inference

Referências

Al-Surayhi, M. M., Al-Qahtani, M. R., Al-Marshadi, S. N., Al-Ghamdi, A. A. Q., e Al-Surayhi, M. M. (2025). Generative artificial intelligence in game design: A narrative review. Journal of Ecohumanism, 4(4):1702–1712.

Bengesi, S., El-Sayed, H., Sarker, M. K., Houkpati, Y., Irungu, J., e Oladunni, T. (2024). Advancements in generative ai: A comprehensive review of gans, gpt, autoencoders, diffusion model, and transformers. IEEE Access, 12:69812–69837.

Dahlström, E., Ferraiolo, J., e Jackson, D. (2011). Scalable vector graphics (svg) 1.1 specification. Technical report, World Wide Web Consortium (W3C).

Dos Santos, I. T. F. (2026). Pixelcraft ai: Svg pixel art generator. [link].

Dos Santos, I. T. F., Barboza, R. S., Almeida, F. P., Araujo, C. S., e Silva Junior, J. M. (2026). Impact of GPU architecture and VRAM on quantized LLM inference for code deobfuscation. In International Conference on Artificial Intelligence, Computer, Data Sciences and Applications (ACDSA). IEEE.

Fukaya, K., Daylamani-Zad, D., e Agius, H. (2024). Evaluation metrics for intelligent generation of graphical game assets. IEEE Transactions on Pattern Analysis and Machine Intelligence.

Google DeepMind (2026a). Gemini 3: Frontier multimodal capabilities. Technical report, Google LLC.

Google DeepMind (2026b). Gemma: Open models based on gemini research. Technical report, Google LLC.

Midjourney Team (2026). Midjourney v6: Generative art and text-to-image models. Accessed: 2026-04-09.

Minini, P. e Assunção, J. (2020). Combining constructive procedural dungeon generation methods with wavefunctioncollapse in top-down 2d games. In Proceedings of SBGames, pages 27–29.

Omena, M. (2025). Indústria de games no brasil vira o jogo no cenário global e com apoio das grandes marcas.

OpenAI (2025). Generative ai in game development and asset creation: Api documentation and guidelines. Technical report, OpenAI. Accessed: 2025-11-09.

Qwen Team (2024). Qwen 3.5: Technical report on large language models. Technical report, Alibaba Cloud Intelligence.

Rogers, S. (2014). Level Up! The Guide to Great Video Game Design. John Wiley & Sons.

Shaker, N., Togelius, J., e Nelson, M. J. (2016). Procedural Content Generation in Games. Springer International Publishing, Cham.

Shojaee, P., Mirzadeh, I., Alizadeh, K., Horton, M., Bengio, S., e Farajtabar, M. (2025). The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity.

Singh, A., Singh, H., e Mattoo, A. (2026). Advancements in generative ai based on large language models and diffusion/gan-based image generation. In Doctoral Symposium on Computational Intelligence, pages 193–206. Springer.

Stability AI (2023). Stable diffusion xl 1.0: Technical report. Technical report, Stability AI.

Summerville, A., Snodgrass, S., Guzdial, M., Holmgård, C., Hoover, A. K., Isaksen, A., Nealen, A., e Togelius, J. (2018). Procedural content generation via machine learning (pcgml). arXiv preprint arXiv:1702.00539.

Werneck, M. e Clua, E. W. (2020). Generating procedural dungeons using machine learning methods. In Brazilian Symposium on Computer Games and Digital Entertainment (SBGames), pages 90–96. IEEE.

Zhang, X., Zhang, Y., Chen, Z., Yu, J., Yang, W., e Song, Z. (2026). Logical phase transitions: Understanding collapse in llm logical reasoning. In Annual Meeting of the Association for Computational Linguistics (ACL). Association for Computational Linguistics.

Zhipu AI Team (2026). Glm-4.7. Accessed: 2026-01-09.