Contribution
Gradient guidance — nudging a diffusion sampler toward an external objective — is standard and is usually implemented by adding the gradient after the denoising step. This says that is the wrong order when the data live on a structured set, and gives the reason: the denoiser is a learned approximate projection, so putting the gradient inside it turns the sampler into an inexact projected-gradient method instead of an unconstrained one with a prior loosely attached.
What is true afterwards: the standard recipe has a named failure mode (off-geometry drift), a one-line fix at identical cost, and convergence guarantees in three geometries where previously the analysis covered linear inverse problems with quadratic objectives.
Key insight
The denoiser is already the projection you were about to hand-write.
D(x) = E[x₀ | x] maps a noisy point toward the data support. If the clean
data lie on a manifold or inside a convex set, that is approximately a
nearest-point projection onto it — and a sampler already calls it once per
step. So a constrained-optimization algorithm is available for free: take the
gradient step in ambient space, then let the denoiser put you back.
The reason PDG fails is the same fact read the other way. Once you have denoised, you are on the geometry; adding a gradient afterwards is precisely the step that leaves it, with nothing following to correct the departure.
Assumptions
- An exact Stein posterior-mean denoiser. The theory assumes the denoiser is the true conditional mean, not a learned approximation of it.
- The learned support faithfully represents the feasible set. The paper says explicitly that projection-error terms can absorb learned-score error and that a full treatment of misspecification is left to future work.
- Deterministic DDIM dynamics, variance-exploding diffusion, an exponentially decaying noise schedule. Other samplers and schedules “require tracking their time-varying relaxation and denoising errors” — i.e. the theory does not transfer as written.
- Compactness in the convex and manifold settings; smoothness of the objective for the nonconvex rates.
Key results
- Bounds between the Stein denoiser and the relevant projection in three settings: linear-Gaussian, compact convex, compact smooth submanifold.
- Finite-time guarantees per geometry: geometric contraction of the projected objective error for strongly convex objectives on linear-Gaussian and convex sets; a best-iterate rate for smooth nonconvex; the same rate locally on submanifolds.
- Cost: one gradient evaluation and one pretrained-denoiser call per reverse step, on a single trajectory — against baselines needing score Jacobians or repeated batch generation.
- Trajectory tracking (double integrator, unicycle): PDG plans track the reference closely and lose much of that advantage on rollout through the true dynamics; DCG’s plans and rollouts nearly coincide. DCG reports the lowest rollout cost and dynamic-feasibility error among guided methods on both random references; on the circle reference its rollout cost is slightly worse than the DRGD baseline while its feasibility error is best.
- D4RL/MuJoCo (Hopper, Walker, HalfCheetah medium-v2): “slightly higher” normalized returns than PDG on all three, guidance scale tuned separately per method.
Claims
Well supported and the reason to file this: that the ordering matters and that the denoiser-as-projection reading explains why. The theory is proved rather than gestured at, in three geometries rather than one, and the mechanism predicts the experimental failure mode of the alternative.
The most useful empirical finding, and it is a measurement finding: PDG’s advantage is largely on the planned trajectory and largely gone after execution. That is not a statement about optimizer quality — it is a statement that the quantity being optimized was not the quantity that mattered. A practitioner comparing guidance methods on planned cost would have chosen PDG.
Thin, and labelled thin by its authors: the reinforcement-learning result. “Slightly higher” on three MuJoCo tasks against one baseline is consistent with the claim and does not carry it, and the paper names broader evaluation as future work.
Proved under assumptions that are not the deployment conditions. Exact posterior-mean denoiser, faithful support, DDIM, variance-exploding, exponential schedule. Every one of those is stated; together they mean the guarantees describe an idealization of the thing people would run.
Not established here: magnitudes. The comparison tables did not survive text extraction from the HTML rendering, so this reading quotes directions and the authors’ own qualifiers rather than effect sizes. Anyone relying on how much better DCG is should read Tables 2 and 3 in the source.
Method
Define DCG as gradient-then-denoise; prove denoiser-to-projection bounds in three geometries; read the reverse process as time-varying inexact projected gradient and derive finite-time rates; evaluate on synthetic constrained problems, double-integrator and unicycle trajectory tracking with rollout through true dynamics, and a Diffuser-style D4RL pipeline.
Connections
- SOTA-203 — sample with a higher-order ODE solver on weights you already trained — is the closest neighbour in kind: an inference-time change to the sampler, no retraining, easy to get wrong in implementation.
- SOTA-266 and SOTA-187 are about how the model is trained and where it lives; this is about what you do to it at sampling time when there is an external objective, which the record did not cover.
- SOTA-235 — update weights at inference on a structurally novel test instance — is the other end of the same spectrum: both adapt at inference, one by changing the weights and this one by changing the update order.
Bearing on the record
One practice and one theory. The practice is the ordering and its conditions;
the theory is the denoiser-as-projection reading, which is the justification
and is Proposed because its proofs assume an exact denoiser and faithful
support.
It also fills a gap in the diffusion line: the record had training-time and sampling-time practices but nothing about steering a pretrained sampler toward an objective, which is the interface between generative models and planning.
Limitations
The guarantees are for an idealization. Exact denoiser, faithful support, one sampler family, one noise schedule. Useful as an explanation of why the ordering helps; not a promise about a learned model.
One group, one paper, and the strongest experiments are in the appendix.
The RL evidence is the weakest part and is the part most likely to be cited, since D4RL numbers travel further than a unicycle tracking study.
Effect sizes unverified in this reading — see Claims.
Open questions
- How much of the guarantee survives a learned denoiser? The authors name misspecification as future work and it is the whole distance between the theorem and the practice.
- Does the ordering still help under stochastic samplers and variance-preserving schedules, which is what most deployed systems use?
- Is the planned-versus-executed gap specific to guidance, or does it show up wherever a diffusion planner is scored on its own output rather than on the consequence of acting on it?