Researchers have built a guidance framework for diffusion models that guarantees constraint satisfaction — no exceptions, no near-misses.
The paper, posted to arXiv, addresses a gap in how conditional generation works today. Existing methods — soft guidance, reward-based steering — nudge a model toward a desired output but can't promise it arrives there. The new approach draws on Doob's h-transform, a tool from probability theory, to inject an explicit drift correction into a pretrained diffusion model's sampling process. Crucially, it does this without retraining the underlying score network, which means it can be bolted onto existing models. Two off-policy learning algorithms estimate the required functions from trajectories the pretrained model already produces.
This matters most in safety-critical settings where "usually works" isn't good enough — think rare-event simulation in financial risk modeling or physical systems where an out-of-bounds sample isn't just wrong, it's dangerous. The authors back the approach with non-asymptotic error bounds in both total variation and Wasserstein distances, which puts it on firmer theoretical ground than most guidance methods in circulation.
Diffusion models have dominated generative AI for years, but their conditioning story has always leaned on probabilistic encouragement rather than hard guarantees. Whether this framework scales cleanly beyond the numerical experiments shown — finance simulations, not image generation — remains the open question.