Here's what actually issued. On July 21, 2020, Adobe Inc. was granted US10719742B2, "Image composites using a generative adversarial neural network," with inventors including Elya Shechtman and Oliver Wang — names that recur across Adobe's research output on image synthesis. The CPC list runs G06K 9/66, G06N 3/04, and G06N 3/088 (unsupervised network training), which tells you immediately this is a vision patent built on a generative adversarial network.
The mechanism, in plain terms: a GAN pits a generator that proposes composited images against a discriminator that tries to spot the fakes, and the two train against each other until the generator's blends look convincing. Applied to compositing — dropping a person or object into a new background — the network learns to harmonize lighting, edges, and color so the inserted element does not look pasted in. For a company whose business is creative tooling, that is a directly monetizable capability.
This is the GAN era of AI patenting, before diffusion models displaced adversarial networks as the default for image generation. Reading a 2020 grant like this one is a reminder that the patent record lags the research frontier: by the time a GAN method issues as an enforceable grant, the field has often moved on to the next architecture. The claim is no less valid for that — it just reflects where the science was when the application was filed.
On scope, the discipline holds. Granted B2, enforceable, but the claims describe a particular compositing pipeline using a GAN. They do not cover every way to merge images, nor the GAN concept itself, which was already widely published. The boundary is in the steps the examiner allowed.
The takeaway: US10719742B2 shows Adobe doing what it does across its portfolio — patenting the specific generative technique that powers a creative feature, naming the same core researchers, and keeping the claims tied to a concrete method rather than a sweeping idea.