US 2026/0183659 A1 · REAL TIME GENERATIVE REMIX

Abstract

Systems and methods for real-time generative remix of rendered content using machine learning models are provided herein. In some embodiments, diffusion-based generative models may be employed to transform game assets and rendered frames in real-time, enabling classic games to adopt the visual characteristics of modern games without requiring offline preprocessing. The disclosed techniques may include feedback mechanisms that analyze previously rendered and generated content to identify and correct visual aberrations, maintaining temporal consistency while improving output quality. In some aspects, the systems may operate at both asset-level generation, where individual game components are transformed during rendering, and screen-space generation, where complete rendered frames are enhanced for display.

Claims20

Dependent claims

A graphics processor comprising:

a memory interface;
and
a graphics processor cluster including a plurality of multiprocessors interconnected via a data crossbar, wherein a multiprocessor of the plurality of multiprocessors is configured to:
receive a command buffer associated with a first accelerated application, the command buffer configured to reference a plurality of assets within a memory accessible via the memory interface;
access a first asset of the plurality of assets in the memory;
generate a second asset based on the first asset via operations performed on behalf of a first machine learning model, the second asset having an appearance of an asset associated with a second accelerated application;
and
process commands of the command buffer to render at least a portion of a scene on behalf of the first accelerated application, the scene rendered based at least in part on the second asset.

5 more elements

Dependencies are inferred from the claim text.

Real time generative remix

Life of this patent

Real time generative remix

Life of this patent

Abstract

Claims20