A new framework called BIRD (Bayesian information restricted diffusion) locates a precise boundary between copying and creating in diffusion models, the systems behind Stable Diffusion and Midjourney.
AI image generators sometimes hand back a near-copy of a training photo and sometimes produce something they have clearly never seen. A new theoretical paper claims the difference is a precise mathematical line, and that it can be located exactly.
The paper introduces a framework called BIRD, short for Bayesian information restricted diffusion. BIRD treats each pixel of a generated image as a deliberately limited observer of the noisy training data the model was fed. Generation, in this view, is the model reasoning about which past training sample could plausibly have produced the noisy pixel it is looking at right now. The smaller the slice of information each pixel can see, the more the model has to guess. The larger the slice, the more it can lean on memory.
The authors derive an inequality that lives in a three-dimensional space defined by three knobs: the amount of training data the model has seen, the timestep of the reverse diffusion process, and the amount of pixel-level information restriction applied. Below the boundary the model memorizes; above it the model generalizes. Move any one of the three knobs and the model can flip between the two. The full derivation is in the preprint.
The framework is exact under minimal assumptions about the data distribution. It is a generalization of earlier analytical work on diffusion models, replacing a fixed local-information assumption with a tunable per-pixel parameter. To check whether the math is grounded, the authors tested whether spatially local BIRD models behave like real, trained diffusion models in the early phase of training. Across two of the dominant image-generator architectures, UNet-based systems (the design behind Stable Diffusion) and DiT-based systems (a newer transformer-based design), the BIRD approximation tracked the real models closely during the first stretch of training. The match held for both architectures, suggesting the boundary is a property of diffusion learning itself rather than of any particular network family.
The paper is a preprint posted in the July 2026 arXiv batch and has not been peer-reviewed. The BIRD approximation is shown to be accurate only in the early-training regime and on the modest datasets used for the validation experiments. The authors are careful not to claim that their boundary settles the legal or practical debate over AI memorization, training-data leakage, or copyright. The line they draw is a theoretical one inside a much larger open problem.
The paper does add one thing worth keeping: a portable mental model. Anyone who has wondered why a model will sometimes regurgitate a near-duplicate of a famous painting and sometimes synthesize a coherent new scene now has a way to talk about the two regimes without appealing to vibes. Below the boundary, the math says, copying. Above it, creation. The next test is whether the same boundary holds once models are scaled to the billion-image corpora behind the public image generators.