Textures provide critical surface details, but high-resolution textures for multiple material properties (e.g., diffuse color, normal maps) create a storage and performance bottleneck. Block Compression (BC) [1, 2, 3] formats are commonly used to reduce texture size, but they are not suited to handle high-dimensional data and don't scale well with increasing resolution.

Recent work [4, 5] has shown the potential of neural networks to model and represent material properties. This neural approach aims at replacing traditional textures with a collection of learned latent features, also known as neural textures, alongside a neural network. In this context, the network plays a crucial role in decoding the learned information and reconstructing the original material

Vaidyanathan et al. [6] have demonstrated how to use these neural representations to improve the compression rate of high-resolution material textures. However, their technique is not suited for real-time applications due to its reliance on a large decoder network and, most importantly, performing post-inference filtering.

At Ubisoft, we have developed a neural representation [7] for material textures that addresses these challenges. Our method uses a lightweight neural decoder with minimal computational overhead, enabling faster inference on GPUs. Additionally, we integrate filtering directly into the neural process, reducing extra computation. While our approach delivers scalable, high-quality texture compression and outperforms traditional BC methods in real-time environments, it currently supports only simple materials with a single texture set.

Extending this technique to efficiently handle complex materials with multiple texture sets is the focus of this internship, paving the way for broader adoption.

References

[1] https://en.wikipedia.org/wiki/S3_Texture_Compression

[2] https://www.reedbeta.com/blog/understanding-bcn-texture-compression-formats/

[3] Campbell, Graham, et al., “Two bit/pixel full color encoding.” ACM SIGGRAPH (1986).

[4] T. Zeltner et al., “Real-Time Neural Appearance Models”. ACM SIGGRAPH (2024).

[5] A. Kuznetsov et al., “Rendering Neural Materials on Curved Surfaces”. ACM SIGGRAPH (2022).

[6] K. Vaidyanathan, et al., “Random-Access Neural Compression of Material Textures”, ACM SIGGRAPH (2023).

[7] C. Weinreich, et al., “Real-Time Neural Materials using Block-Compressed Features”. Eurographics (2024).

• Currently a second-year master’s student or a third-year engineering student.
• Solid foundation in Machine Learning, linear algebra, and signal processing.
• Knowledge of computer graphics fundamentals, such as texture sampling/filtering and shading is a plus.
• Proficiency in Python, and familiar deep learning frameworks (e.g., PyTorch, TensorFlow).
• Proficient in English, both written and spoken, with the ability to clearly communicate technical concepts and collaborate effectively with an international team.

Skills and competencies show up in different forms and can be based on different experiences, that's why we strongly encourage you to apply even though you may not have all the requirements listed above.

This job is open for an internship (6-month contract).

Remote: hybrid model

Process:

• Interview with our recruiter
• One or more technical and project interviews with the manager and his team

If your application is not retained, you will receive a negative answer.

We are working to enrich players’ lives through unique and memorable gaming experiences and by improving the positive impacts of our games. To get there, we are creating a safer, more inclusive work environment, we are giving back to the communities where Ubisoft operates by working with local non-profit partners and by working to reduce the environmental impact of our business. Learn more on our Social Impact here.

Check out this guide to help you with your application, and learn about our actions to encourage more diversity and inclusion.