LIT-490: Denoising as Projection: Constrained Optimization with Gradient-Guided Diffusion

Zhang, Zhang, Zardini, Amin and Ozdaglar (2026) — ARXIV-2608.29507

Key takeaways

• The intervention is an ordering, and that is the whole appeal. Standard gradient guidance — the paper calls it post-denoising gradient guidance, PDG — computes the denoised estimate and then adds −η∇f outside it. DCG applies the gradient to the noisy iterate first, then passes the result through the denoiser. Same two ingredients, same cost, opposite order.
• Why the order matters is geometric. When the data lie on a structured feasible set — a manifold, a convex region, a family of dynamically admissible trajectories — the objective gradient need not be tangent to it. Added after denoising, it moves the sample off the geometry the model learned. Added before, the denoiser gets a chance to pull it back, because the Stein posterior-mean denoiser maps noisy points toward the data support.
• That reading is formalised, not asserted. The paper proves bounds between the Stein denoiser and the relevant projection in three settings — linear-Gaussian, compact convex, compact smooth submanifold — and then reads the reverse process as a time-varying inexact projected-gradient method, with finite-time guarantees in each: geometric contraction for strongly convex objectives, a best-iterate rate for smooth nonconvex.
• The cost claim is the selling point against the alternatives. One gradient evaluation and one denoiser call per reverse step, a single trajectory. The comparison table has the nearest optimization-oriented methods requiring score Jacobians or multiple reverse trajectories; manifold-aware guidance and reward alignment generally need fine-tuning or auxiliary models.
• The planning experiments are the informative ones, and they are a measurement story. In double-integrator and unicycle trajectory tracking, PDG produces plans that follow the reference closely — and much of that gain disappears once the controls are executed through the true dynamics. DCG’s planned improvement survives rollout: it reports the lowest rollout cost and dynamic-feasibility error among guided methods on both random references. The finding is less “better optimizer” than “PDG was optimizing the plan rather than the outcome”.
• The reinforcement-learning result is thin and the paper says so. On Hopper, Walker and HalfCheetah (medium-v2) under a Diffuser-style pipeline, DCG achieves “slightly higher” normalized returns than PDG on all three. Three tasks, one baseline, a qualifier the authors chose themselves.

Standing in the anthology

The record’s diffusion line is about how to train and how to sample — SOTA-266 on the interpolant and timestep distribution, SOTA-203 on using a higher-order solver with weights you already have, SOTA-187 on working in a learned latent. This adds the case where sampling is being bent toward an external objective, which none of them covers and which is where diffusion models meet planning and control.

It belongs beside SOTA-203 in kind as much as in subject: both are inference-time changes that need no retraining, and both are the sort of one-line edit that is easy to get backwards.