OYS Directly Searches Diffusion Model Denoising Timesteps, Preserving 89%–94% of 50-Step Quality in 5 Steps

Diffusion model inference cost is typically measured by the number of denoising-network forward passes. Engineering implementations often simply shorten the default timestep schedule, but this may not preserve the best possible quality at low step counts. Researchers from Cornell University and other institutions have introduced Optimize Your Sampling (OYS), which treats sampling parameters—including timesteps and guidance strength—as a black box and directly uses Bayesian optimization to search for configurations that improve a target metric. It does not require backpropagation through the full sampling chain or a differentiable scoring function, allowing it to optimize HPS, LPIPS, PSNR, or task-specific non-differentiable scores.

This differs from Align Your Steps (AYS), which optimizes an upper bound on the KL divergence arising from diffusion SDE discretization. In each iteration, OYS generates a batch of images, after which a surrogate model and acquisition function select the next candidate timestep configuration. This approach is more expensive, but it directly optimizes the final quality metric used in practice. The search cost is paid only when tuning each model and task, and can subsequently be amortized over a large volume of production inference.

In five-step tests on COCO Captions, OYS achieved HPS scores of 0.252, 0.238, and 0.245 for DeepFloyd, Stable Diffusion 1.5, and SDXL, respectively. Their default schedules at the same step count scored 0.215, 0.230, and 0.229. For SDXL, OYS images achieved a 70.70% HPS win rate over images generated with the default schedule. Even with SDXL-Turbo, which was already distilled for few-step generation, three-step OYS increased the score from 0.287 to 0.298. The five-step win rates for FLUX.1-dev and Qwen Image were 82.69% and 77.19%, respectively. An SDXL study involving 58 participants also found that users preferred OYS for both fidelity and text alignment.

The study summarizes the five-step schedules as preserving 89%–94% of the quality achieved with 50 steps, suggesting that mature models can reduce neural-network evaluations by roughly 10× without retraining their weights. However, this does not translate into a 10× end-to-end serving speedup: text encoding, decoding, batch scheduling, and data transfer still impose fixed costs. The optimal schedule also depends heavily on the model, prompt distribution, sampler, and evaluation metric. Tuning FLUX.1-dev, for example, required generating 5,376 images. The next questions for deployers are whether these schedules generalize across LoRA adapters, quantized variants, and real-world Chinese-language prompts, and whether the authors will release reproducible search code and schedule files.