We formalize this idea with a definition called Projective Composition — based on projection functions that extract the “key features” for each distribution to be composed. 4/

Part of challenge is, we may want compositions to be OOD w.r.t. the distributions being composed. For example in this CLEVR experiment, we trained diffusion models on images of a *single* object conditioned on location, and composed them to generate images of *multiple* objects. 2/

Paper🧵 (cross-posted at X): When does composition of diffusion models "work"? Intuitively, the reason dog+hat works and dog+horse doesn’t has something to do with independence between the concepts being composed. The tricky part is to formalize exactly what this means. 1/

finally managed to sneak my dog into a paper: arxiv.org/abs/2502.04549

nice idea actually lol: “Periodic cooking of eggs” : www.nature.com/articles/s44...

Just read this, neat paper! I really enjoyed Figure 3 illustrating the basic idea: Suppose you train a diffusion model where the denoiser is restricted to be "local" (each pixel i only depends on its 3x3 neighborhood N(i)). The optimal local denoiser for pixel i is E[ x_0[i] | x_t[ N(i) ] ]...cont

Catch our talk about CFG at the M3L workshop Saturday morning @ Neurips! I’ll also be at the morning poster session, happy to chat