Across the portfolio, 2023 is when diffusion-based image generation floods the record. A PatentBear sweep for text-to-image and diffusion methods in the 2023 window returns thousands of documents — a sharp jump from prior years that mirrors the public explosion of generative imaging. The structural finding: the technique cluster shifted decisively from GANs to diffusion, and the volume of filings tracks the hype curve almost exactly.
A representative exemplar is US20240296596A1, "Personalized Text-to-Image Diffusion Model," with inventors including Nataniel Ruiz and Michael Rubinstein — Google-research names behind personalization work that lets a diffusion model learn a specific subject from a few images. It typifies the 2023-24 wave: diffusion at the core, personalization and control as the differentiators.
The CPC clustering is informative. These records concentrate in image-generation code G06T 11/00 and vision classes G06V 10/82 and G06V 10/764, with G06N 3/08 for the learning. That footprint distinguishes them from the older GAN cluster and marks the architectural changeover. The named-assignee leaders among the substantive AI filings skew toward Adobe and Google research, but — and this is the key caveat — the single largest bucket is unassigned.
Two caveats bite hard in a hype-driven year. First, the raw count is inflated by the breadth of the query and by adjacent fields (semiconductor and medical-imaging records share some codes), so the headline number overstates the pure generative-AI subset. Second, these are overwhelmingly applications, not grants — the 2023 surge is a filing surge, and how much of it issues, and with what scope, won't be clear for years.
For the strategist: read the 2023 diffusion record as a leading indicator of where the labs are racing, not as a settled ownership map. Anchor any claim to specific records like US20240296596A1, weight grants over applications, and discount the aggregate counts for query breadth and prosecution lag.