Optimizing high‑dimensional physical systems is challenging due to expensive evaluations and complex constraints. In offline model‑based optimization (MBO) only a fixed dataset is available and the true objective cannot be queried at test time. Existing approaches tend to be conservative, over‑exploit unconstrained surrogates, or act as static samplers that cannot refine candidates against precise physical objectives.\ \ We introduce Differentiable Diffusion Conditioning (D2C), which exploits the fact that deterministic sampling renders the reverse denoising chain differentiable with respect to the conditioning signal. By freezing a conditional diffusion model, D2C turns it into an end‑to‑end gradient‑based optimizer: conditioning variables are optimized at test time using gradients from frozen surrogate objectives and physics‑informed penalties back‑propagated through the reverse chain. The method is realized within a proposer‑evaluator framework and evaluated on laser‑pulse optimization for inertial confinement fusion, sustainable data‑center workload scheduling, and four Design‑Bench benchmarks. Results show the strongest performance on the two real‑world tasks and the best average rank across the Design‑Bench suite.\ \ Review