Nvidia filed 2 patents, all in AI & Machine Learning.
The filings cover methods for enhancing conversational AI by enabling chatbots and NPCs to insert relevant images into text responses without relying on resource-intensive multimodal models. Additionally, the AI & Machine Learning patents address auto-regressive techniques for generating detailed 3D meshes with thousands of faces in specific artistic styles, supporting efficient creation of game assets and virtual environments through mesh tokenization.
The pair of AI & Machine Learning patents tackle two distinct challenges in content generation. The first addresses the computational expense of multimodal AI by separating image retrieval from language processing. Rather than training massive models that handle both text and images, the approach pre-indexes text-to-image relationships from document collections and uses vector similarity matching to find relevant visuals for chatbot responses, delivering multimedia conversations at a fraction of the compute cost. The second patent generates complex 3D meshes with over 8000 faces through an auto-regressive auto-encoder architecture that compresses mesh data into compact latent representations. This tokenization method produces meshes with 5 times more faces than previous techniques while preserving topology that meets artist quality standards, and the system supports varied conditioning inputs to generalize beyond its training data for creating game assets and VR content.
All data sourced from USPTO patent filings. Google Patents may take several weeks to index recent publications. If a link is unavailable, search for the patent number at USPTO Patent Public Search.