Here's what actually issued. On August 18, 2020, Microsoft Technology Licensing, LLC was granted US10748314B2, "Controllable conditional image generation," with inventors Gang Hua and Navaneeth Kumar Bodla. The CPC list — G06T 11/60, G06T 11/00, G06N 3/08 — places this in image generation backed by a trained network.
The idea behind "conditional" generation is control. An unconditional generator produces a plausible image from noise; a conditional generator takes an additional input — a label, an attribute, a layout — and produces an image that satisfies it. The "controllable" framing in the title points at letting a user steer attributes of the output rather than accepting whatever the model emits. That is the difference between a novelty demo and a usable creative tool.
For Microsoft, sitting on both research and a productization pipeline across Office and Azure, owning methods for steerable generation is strategically sensible. The 2020 date again places this in the pre-diffusion generative era, when controllable GAN-style generation was an active research front.
On scope: granted B2, enforceable, but the claims describe a specific method for conditioning and controlling generated images. The terms "controllable" and "conditional" in the title are descriptive, not the legal boundary — claim 1 is. Read it before describing what Microsoft can assert.
The takeaway: US10748314B2 is another data point in the major labs' steady accumulation of generative-imaging IP years before generative AI became a consumer phenomenon. The claims are method-specific, the assignee is a hyperscaler, and the filing reflects the architecture of its moment.