Here's what actually issued. On September 3, 2024, Lemon Inc. was granted US12079902B2, "Generation of images corresponding to input text using multi-algorithm diffusion sampling," inventors Qing Yan, Bingchen Liu, Yizhe Zhu, and Xiao Yang. The CPC codes are image-generation G06T 11/00, NLP G06F 40/40, and G06T 5/70 (denoising).
The mechanism is the sampler. A diffusion model generates an image by starting from noise and iteratively denoising it toward a clean picture; the sampling algorithm governs how that denoising trajectory is taken. Different samplers trade off speed and quality. "Multi-algorithm" sampling means combining or switching among multiple sampling algorithms during generation — using the strengths of each — to produce text-conditioned images more efficiently or with better fidelity than any single sampler.
This grant is one of the first wave of diffusion-specific patents to actually issue, distinguishing it from the larger pile of pending diffusion applications. The denoising CPC (G06T 5/70) underscores that the claimed novelty is in the generative sampling process itself, not just the text-conditioning. As text-to-image moves from research to product, the sampling layer — where compute cost is decided — becomes a competitive surface worth owning.
On scope: granted B2, enforceable, but the claims describe a specific multi-algorithm diffusion-sampling method for text-to-image. They do not cover diffusion models, samplers, or text-to-image generally. Claim 1 defines the line.
The takeaway: US12079902B2 is a granted diffusion patent — not just an application — pinpointing the sampling process as the inventive step, a sign that the diffusion-era IP is starting to mature from filings into enforceable rights.