PEEK uses a look-ahead prediction to improve diffusion image sampling

Diffusion models learn to denoise corrupted examples of real images. During generation, however, they work on states created by their own previous predictions. Yunwei Bai, Ying Kiat Tan and Tsuhan Chen, of the National University of Singapore, describe this mismatch as exposure bias and propose PEEK to reduce its effects without retraining the generator. Their paper is available as an arXiv preprint; the supplied landing page reports acceptance to NeurIPS 2026.

The problem is especially relevant when generation uses few denoiser evaluations. Large steps must then carry an initially noisy state toward a finished image, leaving less opportunity to recover from imperfect predictions. The authors argue that spending some computation on correcting these predictions can be more useful than spending it entirely on additional ordinary sampling steps.

Look ahead, then revise the current step

PEEK first obtains the usual estimate of the clean image from the current noisy state. At a selected position, it makes a provisional update to the next, lower-noise timestep and queries the denoiser there. It then returns to the original state and constructs the actual update using the look-ahead clean prediction.

Although the paper calls this interpolation, the main PEEK method uses binary gates: either retain the current prediction or replace it with the look-ahead prediction. The final denoising step cannot use this correction because no later noisy timestep remains within the trajectory. For latent diffusion models, the operation modifies the predicted clean latent representation, not a decoded image.

The rationale is that the provisional state has a higher signal-to-noise ratio and may support a better clean prediction. That is not guaranteed. The paper explicitly recognizes that an additional look-ahead query can reinforce bias. Its mathematical argument says that reducing clean-prediction error tightens an upper bound on the corresponding one-step state error; it does not establish that every look-ahead prediction reduces that error.

Choosing where the extra calls help

PEEK selects correction positions through a greedy validation search. A correction is accepted only if it improves the target metric over both the current partial schedule and an ordinary baseline using the proposed correction’s denoiser-evaluation budget. The experiments use Fréchet Inception Distance, or FID, where lower values are better.

This accounting matters. A trajectory with N ordinary denoiser calls and K active corrections costs N plus K calls. The comparison baseline therefore receives N plus K ordinary evaluations, rather than just N. PEEK is not free relative to its shorter starting trajectory; the reported benefit is a better use of a matched evaluation budget.

Calibration is specific to a model, sampler, step count and target metric. Once selected, the schedule stays fixed during deployment. Two caches reduce repeated search work: one reuses selected PEEK prefixes, while the other approximately rebases ordinary baseline states onto changed timestep grids. The authors’ worst-case nine-step accounting falls from 225 to 78 denoiser queries, a 65.3% reduction. Baseline rebasing is approximate because differing step intervals can introduce mismatch.

Results across image-generation settings

The experiments cover CIFAR-10, LSUN Bedroom, ImageNet and MS-COCO, including pixel-space and latent-space generators and unconditional, class-conditioned and text-conditioned sampling. Schedules selected using lightweight proxy metrics are subsequently evaluated with fixed settings; final FID samples are distinct from validation-search samples.

The largest reported relative FID reduction is 49.3% on the 64-by-64 ImageNet EDM model with DPM-Solver++: FID falls from 19.71 to 9.99. That PEEK configuration uses six base calls and three look-ahead calls, matching a nine-evaluation baseline. On the basic LSUN Bedroom model with DEIS, FID falls from 20.57 to 11.38, a 44.7% reduction.

Gains are not uniform. On the 256-by-256 latent ImageNet model, FID changes from 20.40 to 19.64. For first-order text-conditioned MS-COCO generation, it falls from 14.58 to 12.93, an 11.3% reduction. These percentages describe FID changes, not equivalent percentage increases in perceived image quality.

A variant, PEEK-G, blends solver directions rather than clean predictions and allows fractional gates. Its reported FID reductions include 24.6% on CIFAR-10 Model B and 26.0% on MS-COCO, but much smaller changes on basic LSUN and ImageNet.

Costs and qualifications

Reported calibration takes at most 2.2 minutes in these experiments, run on multi-GPU H200 nodes; the MS-COCO proxy search reports 102.2 GB peak CUDA memory. Those timings are hardware-specific, not general deployment promises.

The paper also contains an unresolved coverage qualification: its conclusion names reduced coverage as a limitation, while the main basic-LSUN table reports coverage increases against the corresponding matched baselines. Together with the model-dependent gains and the risk of biased look-ahead predictions, this supports a bounded conclusion: PEEK improves the reported sampling comparisons, rather than guaranteeing better or more diverse images in every setting.