Graphics & Rendering

Ubisoft Entertainment received 1 patent for texture compression powered by neural networks. The system trains a per-material network to handle both decompression and trilinear filtering in fewer than 200 GPU operations, storing network weights as 16-bit floats while applying traditional block compression formats like BC6 or ASTC to the neural features themselves. Training happens end-to-end via backpropagation against a perceptual loss function, which allows the network to learn and counteract compression artifacts in ways that fixed mathematical transformations cannot. The approach targets real-time 4K rendering while dramatically cutting VRAM requirements compared to conventional PBR material storage.

Intel filed 2 patents addressing real-time visual enhancement during gameplay. The first applies generative AI models on-the-fly to transform legacy game graphics without requiring offline asset processing or additional storage, enabling aesthetic changes that extend beyond simple resolution upscaling. The second detects stutter conditions proactively by monitoring block-level render timing during the current frame's rendering process, then configures interpolated frame distances adaptively rather than relying on static presets. Both applications focus on improving visual output during active play, whether through dynamic style alteration or responsive frame smoothing.

Qualcomm received 1 patent for a rendering pipeline that shifts post-processing work from the GPU to a dedicated neural processing unit. The architecture sends GPU-rendered frames to the NPU at the same frame rate for real-time neural enhancement, allowing the NPU to handle inference-based image improvements without stalling the GPU's render pipeline. This formalizes a heterogeneous compute model where visual quality enhancements run on specialized hardware, preserving GPU performance budget for core rendering tasks.

Adobe filed 1 patent covering machine learning-based polygon mesh upscaling combined with progressive refinement. The system transmits a coarse mesh first, then uses a trained ML model to upscale it for immediate rendering at higher fidelity before full-quality residual data arrives. The ML model trains on the originating mesh dataset, enabling domain-specific reconstruction tailored to the source geometry rather than generic upsampling. This reduces initial data transfer requirements while preserving eventual visual accuracy through the residual correction layer.

Nvidia received 1 patent for multi-frame interpolation using deep neural networks in real-time gaming and VR contexts. The approach generates multiple interpolated frames per rendered frame while maintaining the low-latency requirements of competitive and immersive environments, extending beyond typical single-frame interpolation methods. The system targets higher perceived frame rates without adding performance penalties to the underlying render process.

Nintendo filed 1 patent for a voxel game camera system that transitions dynamically between close-follow and wide-area aerial views. The system repositions and reorients the virtual camera based on player-character position and gaze point, enabling expansive world exploration without requiring a separate map interface. Mode-specific level-of-detail rendering and dithering transparency in the pull-back view address occlusion challenges unique to voxel terrain environments, maintaining mesh fidelity across both camera perspectives.