Composition of Pretrained Diffusion Models

Created on February 13, 2026

2026   ·   paper   iclr

Our ICLR 2026 paper, Composition of Pretrained Diffusion Models: A Logic-Based Calculus, studies how to compose pretrained diffusion models using operators grounded in fuzzy logic.

The paper introduces conjunction, disjunction, and negation operators that generalize classical mixtures while improving mode coverage, stability under negation queries, and compositional behavior. It also studies Dombi operators for richer logical compositions and derives Feynman-Kac correctors to reduce sampling bias.

Full write-up coming soon.



Enjoy Reading This Article?

Here are some more articles you might like to read next:

  • Temperature Scaling in Discrete Sequence Models