Diffusion models generate an image by starting from random noise and removing a little of it at each step. A growing body of work spends extra computation at generation time to get better images: it branches the generation into several candidate paths, scores the partial results, and keeps only the promising ones. A partially denoised image is unrecognizable, so these methods score it through a "preview." The preview is the model's estimate of the finished clean image, computed directly or with Tweedie's formula (the posterior mean of the clean image given the noisy one). We argue that this preview answers the wrong question. A search algorithm needs to know how good the finished images reachable from a noisy state will be, on average. The preview reports how good the average finished image looks. Early in generation that average is a blur, and the two quantities come apart.
The proposal replaces the single preview with a few cheap stochastic samples of possible finished images, scores each one with the reward model, and averages the scores. The spread of those scores also gives an error bar. A beam search uses that error bar to decide when a pruning decision is justified, which tells the search how much to trust its scores at each step. The claim to test has two parts. At an equal number of network calls, this sampled estimate tracks the true quality of a noisy state better than either the Tweedie preview or a deterministic multi-step lookahead. It also produces better final images per unit of compute, especially for rewards that score structure the blurry average destroys, such as counting, attribute binding, and collision-free paths.
Core Claim
The quantity a search over noisy states actually needs is the value of a state $x_t$ at noise level $t$:
$$V(x_t) = \mathbb{E}_{x_0 \sim p(x_0 \mid x_t)}\big[r(x_0)\big], \qquad \text{whereas previews compute } r\big(\hat{x}_0(x_t)\big),\ \ \hat{x}_0(x_t)=\mathbb{E}[x_0 \mid x_t].$$
Here $r$ is the reward or verifier score of a finished image and $p(x_0 \mid x_t)$ is the distribution of finished images the model would produce from $x_t$. For nonlinear rewards these differ by a gap that grows with the width of the posterior, and so with noise level.
Diffusion Tree Sampling motivates this. It observes that Tweedie estimates are "essentially random" at high noise and responds by running full rollouts to the end and propagating their rewards back up the tree. That is accurate but expensive, and the authors name a learned value function as the open fix. Our training-free estimator sits between the two extremes:
$$\hat{V}_K(x_t) = \frac{1}{K}\sum_{k=1}^{K} r\big(\tilde{x}_0^{(k)}\big), \qquad s_K^2(x_t)=\text{sample variance of the } r\big(\tilde{x}_0^{(k)}\big).$$
Each $\tilde{x}_0^{(k)}$ comes from a short $m$-step stochastic sampler started at $x_t$ with fresh noise. The model's direct clean-image prediction makes these few-step jumps possible. The cost is $K\cdot m$ network calls plus $K$ reward calls per node.
What is new is threefold. The search targets the mean of rewards, not the reward of the mean. It uses the variance $s_K^2$ to decide when pruning is statistically warranted. And it measures how the reliability of each estimator depends on noise level and on whether the network predicts noise, velocity, or the clean image.
Relation to Prior Work
Four papers supply the search machinery and define the comparison, and all four score states with a single point estimate. Scaling Inference Time Compute for Diffusion Models frames scaling as verifiers plus search algorithms (random, zero-order, and path search) and shows search beats adding denoising steps. Test-Time Scaling via Noise Trajectory Search runs a per-step bandit over injected noise and scores candidates with a Tweedie preview; its non-monotonic scaling curves are what one would expect from a biased early value signal. Inference-time scaling through classical search adds BFS and DFS backtracking with adaptive thresholds and Langevin local search; the threshold sensitivity it reports is the problem our variance-aware pruning addresses. Diffusion Latent Beam Search is the closest work. Its lookahead runs a few deterministic DDIM steps to get a cleaner preview. That is still one point, the end of a single deterministic path, and it carries no uncertainty. We keep its beam structure and change only the estimator, so the comparison is direct.
Several papers that use the preview as a plug-in interface for frozen external models constrain the design. Energy Guided Diffusion, Longtail-Guided Diffusion, and Safe Text Embedding Guidance show that frozen classifiers can be evaluated on $\hat{x}_0$ without retraining on noisy inputs. They also show the two costs of doing so: backpropagating through decoders is expensive, and strong gradients push images toward artifacts. Our estimator is zeroth-order and only needs forward reward calls, which avoids the decoder backward pass. ACDMSR and KinemaDiff show that clean-target networks support short trajectories of 40 and 10 steps. That is the property that makes $m$-step posterior samples cheap.
TTGen lets a 26-billion-parameter vision-language model read previews. More Thinking, Less Seeing? warns that such models drift toward linguistic priors as they reason longer. We therefore keep verifiers small and fixed and do not use model "thinking" as the scaling axis.
Research Program
The first stage is a measurement study, and it has value even if the method fails. For each model, reward, and noise level, we draw a few hundred noisy states. For each state we compute a reference value from 64 full stochastic rollouts. We then report the Spearman rank correlation between the reference and each estimator at matched network calls: the Tweedie preview, DLBS lookahead with $K\cdot m$ deterministic steps, and $\hat{V}_K$ for several $(K, m)$ splits. We also report each estimator's bias.
We will run this on three parameterizations. SiT predicts velocity and is the ImageNet family used by Ma et al. EDM uses a preconditioning closer to clean-image prediction and was used by Noise Trajectory Search. Stable Diffusion and PixArt-Σ are the text-to-image models. The hypothesis is that noise-predicting models produce the worst early previews. The reason is that recovering the clean image from a noise prediction multiplies the network's error by $\sigma_t/\alpha_t$ (noise scale over signal scale), and that ratio is large at high noise. We also expect the gap between the mean-of-rewards and reward-of-mean estimates to be largest for compositional and counting rewards.
The second stage is the search algorithm. It is beam search in the style of DLBS: $B$ beams with $J$ children each. The difference is that a child is discarded only when its upper confidence value, $\hat{V}_K + c\, s_K/\sqrt{K}$, falls below the $B$-th best lower confidence value. Otherwise it survives, with surviving beams subsampled at random if needed. At high noise the error bars are wide, so pruning is automatically lenient and the search behaves like parallel best-of-N. In the middle "demixing" phase that Noise Trajectory Search identified, the error bars tighten and the search becomes selective. This replaces the hand-tuned per-step schedules and DFS thresholds of prior work with a single confidence constant $c$. Everything is batched in parallel, which avoids the sequential latency of tree search.
Baselines
The naive reference point is the base sampler with more denoising steps, and next to it Best-of-N at the same total network calls with selection at the end. Ma et al. show both are weak in absolute terms. For the most informative upper reference, they report that PixArt-Σ with search beats FLUX.1-dev on DrawBench at one tenth of FLUX's compute. Best-of-N matters because it is perfectly parallel and the main contrast with our method is a better use of the same calls.
The strongest existing method is DLBS with its lookahead estimator (Diffusion Latent Beam Search). It reports beating Best-of-64 with a small search budget. We found no reported number on Lacuna for DLBS on image benchmarks, and no exact split scores for DLBS on DEVIL, MSRVTT-test, Movie Gen Video Bench, or TVGE. We will reimplement it on the same models, rewards, and beam sizes, and give it the identical per-node call budget $K\cdot m$ as deterministic lookahead steps. Both methods get the same tuning budget, a grid of equal size over lookahead length and over $c$ respectively.
The second strong competitor is DTS* (Diffusion Tree Sampling), which reports matching Best-of-N on DrawBench with 2–5× less compute and pays for accurate values with full rollouts. We also include Noise Trajectory Search, which reports a 164% classifier-reward gain over naive sampling on EDM ImageNet 64×64, and DFS backtracking from classical search. All methods run at matched total network calls and matched reward calls, and we report wall-clock time on the same GPU.
Two ablations isolate the key ingredients. The first uses $K=1$ stochastic sampling, which isolates averaging from merely adding noise. The second keeps $\hat{V}_K$ but uses plain top-$B$ pruning with no confidence bounds, which isolates the use of variance.
Beating the baselines means strictly higher held-out-verifier reward than DLBS lookahead at every budget on at least two of the three task families, and reaching DTS* quality with fewer network calls.
Why not just…?
Why not just add more denoising steps? Scaling Inference Time Compute for Diffusion Models and Noise Trajectory Search both find that returns plateau after a few dozen steps, because the particular noise realization dominates the outcome. We include step scaling as a curve only to anchor the compute axis.
Why not just run Best-of-N in parallel? Best-of-N wastes compute finishing trajectories that were hopeless early. DTS* reports matching it with 2–5× less compute. There is a real threat, though: if early value estimates are as unreliable as the measurement stage may show, adaptive pruning gains little over Best-of-N. Our confidence rule degrades gracefully toward Best-of-N in that regime rather than below it. The measurement study will report exactly where in the generation, if anywhere, pruning pays.
Why not just use DLBS's deterministic lookahead with more steps? This is the most serious alternative. If the posterior over finished images is narrow at the timesteps where pruning happens, a single deterministic peek may be as good as averaging samples, and our method collapses into it. The matched-call rank-correlation experiment settles this directly, broken down by timestep and by reward type. We predict our advantage concentrates in early timesteps and in counting and binding rewards. If it does not appear there, we will report that and drop the claim.
Why not just train a value network on noisy states? That is the AlphaZero-style route DTS leaves open, and it may be more accurate than our estimator. However, it must be retrained for each reward and each base model, which gives up the training-free, frozen-verifier property that motivated preview-based guidance. We will train one value head for the CLIP reward on PixArt-Σ as an upper bound and report how much of its gain $\hat{V}_K$ recovers without training.
Proposed Experiments
The experiments cover three task families, all taken from the file.
The first is class-conditional ImageNet. We use SiT and EDM at 64×64 and 256×256, with a classifier reward, measuring FID and Inception Score.
The second is text-to-image. We use PixArt-Σ and Stable Diffusion on DrawBench and T2I-CompBench. Search optimizes CLIP or DINO, and evaluation uses B-VQA and UniDet. Optimizing one verifier and evaluating with another follows the double-verifier logic of classical search and exposes reward hacking. For calibration, GORS fine-tuning raised B-VQA attribute binding on Stable Diffusion v2 from 0.5065 to 0.6603 (T2I-CompBench). We ask whether search alone on a frozen model approaches that, and whether the two add together.
The third is PointMaze, where the reward is an exact collision check that cannot be hacked. Classical search reports success rising from near zero to nearly 100% with local search. Our metric there is the number of network calls needed to reach high success.
The decisive checks come in order. The first is the per-timestep rank-correlation curves. If $\hat{V}_K$ does not beat both Tweedie and DLBS lookahead at matched calls in the early-to-middle steps, the method has no basis. The second is the quality-versus-compute curves against DLBS and DTS* on the held-out verifier. The third is the ablations separating averaging from confidence-aware pruning. The fourth is the parameterization comparison, testing whether clean-target or velocity models need smaller $K$ than noise-predicting ones.
References
- Diffusion Tree Sampling: Scalable inference‑time alignment of diffusion models
- Scaling Inference Time Compute for Diffusion Models
- Test-Time Scaling of Diffusion Models via Noise Trajectory Search
- Inference-time scaling of diffusion models through classical search
- Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search
- Energy Guided Diffusion for Generating Neurally Exciting Images
- Generative Data Mining with Longtail-Guided Diffusion
- Training-Free Safe Text Embedding Guidance for Text-to-Image Diffusion Models
- ACDMSR: Accelerated Conditional Diffusion Models for Single Image Super-Resolution
- KinemaDiff: Towards Diffusion for Coherent and Physically Plausible Human Motion Prediction
- TTGen: Incorporating Test-time Scaling to Diffusion Models
- More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models
- T2I-CompBench: A Comprehensive Benchmark for Open-world Compositional Text-to-image Generation