Assassin's Creed Mirage, developed by Ubisoft, introduced neural texture compression, a technique that replaces multiple high-resolution texture files with a single lower-resolution map plus a compact neural network, featuring selective runtime reconstruction of material properties.
Ubisoft described the implementation in a developer blog and referenced demonstrations shown at 2024 GDC, tying the tech directly to Mirage's shipped build. The method compresses texture data by having the engine run a small model to recreate surface details such as reflectivity and roughness from a compact map.
Ubisoft applied the approach selectively to high-instance assets like foliage and furniture to reduce memory pressure without impacting unique hero textures. For players and developers this matters because the approach cut VRAM and storage use by up to about 30%, improving performance on memory-limited systems and signaling a practical AI-driven pipeline for real-time game rendering.
Image Credit: Ubisoft Bordeaux
What Makes This Trend Stand Out
- Neural Texture Compression
- A method that replaces multiple high-resolution textures with a compact map plus a small neural model enables substantial VRAM and storage reductions for real-time rendering.
- Selective Runtime Reconstruction
- This technique reconstructs material properties on demand for high-instance assets, preserving visual fidelity on unique hero assets while lowering memory pressure.
- AI Driven Asset Pipelines
- By integrating small neural models into the rendering pipeline, content creation and delivery workflows can shift toward lighter on-disk representations paired with intelligent runtime synthesis.
Sectors Adopting This
- Video Game Development
- Memory-sensitive platforms and open-world titles benefit from reduced asset footprints, allowing larger worlds or higher frame rates without proportional increases in storage or VRAM.
- Graphics Hardware
- GPU and dedicated VPU designs could be impacted by new workload patterns favoring efficient inference for many small neural models alongside traditional rasterization tasks.
- Film and Visual Effects
- Production pipelines that manage vast texture libraries may see opportunities to store lighter representations and reconstruct high-detail surfaces during compositing or playback.
