These three things are the same story: a designer's request for options, a language model, and a structured document. Adobe's grant US11644961B1 (issued May 9, 2023; with a companion publication US20230305690A1) applies a transformer generative model to produce variations of a design document. Generative AI, aimed not at chat but at layout.
Connect the dots to why this is clever. A design document is structured, text blocks, positions, styles. Treating it as something a language model can read and rewrite means the model can propose variations that respect the structure: same content, different arrangement, or same layout, different copy. The CPC tags G06F 40/106 (document layout) and 40/284 (tokenized text) reflect that document-structure focus.
Follow both the money and the IP. Adobe's entire business is creative tooling, and the generate me options task is the daily grind of design work. Owning the method that automates it is a direct product play, the patent is downstream of a feature millions of creatives would use, not a research curiosity.
This is also a clean example of generative AI escaping the chatbox. The public conversation fixates on text and images, but the same transformer machinery is being pointed at structured artifacts, documents, spreadsheets, designs, where the payoff is automating tedious variation. That quiet diffusion into productivity tools is where a lot of the real value lands.
House caveat: it's a granted patent, so the claims bound the real scope, and generates variations is a capability description, not a quality guarantee. As a dated marker it's sharp, by mid-2023, a major creative-software company had patented using a language model to do the designer's option-generation work.