analiza aiqbench
This research addresses exposure bias in diffusion models by introducing PEEK, a training-free sampler plugin that uses look-ahead evaluation to improve image fidelity without retraining denoisers. The method achieves significant FID improvements across various model architectures while maintaining minimal computational overhead during calibration.
PEEK: One-Step-Look-Ahead Exposure Bias Correction for Diffusion Sampling
arXiv:2611.05407v1 Announce Type: cross Abstract: Diffusion image generators are successful due to accurate denoising networks and strong numerical samplers. Yet the two components are not fully matched. During sampling, the denoiser predicts for the next step based on flawed input that differ much from the real data. This train-test gap, or exposure bias, caps image generation quality (i.e., blurry, lacking in details or awkward in shape). To mitigate the problem, we propose PEEK, a novel training-free deterministic sampler plugin. At each eligible denoising step, PEEK first performs a provisional one-step update, evaluates the denoiser once at the look-ahead state, and interpolates it with the one-step-later prediction to form the next noisy state. By taking the additional step, we obtain a generally more-image-like prediction from the higher-signal noisy state, mitigating the exposure bias. To optimize the interpolation positions, PEEK searches for an interpolation schedule search made efficient by the prefix and baseline noise-schedule-rebasing caches. On 8 pixel-space and latent-space diffusion models with unconditional, label-guided, and text-guided settings across various output resolutions, PEEK consistently improves over NFE-matched baselines, with FID improvements up to 49.3\% at minimal extra computation cost (e.g., the one-time-per-model calibration takes at most 2.2 minutes across our experiments.). Project website is at https://dachengruoque.github.io/PEEK-Video/.
Abstract:Diffusion image generators are successful due to accurate denoising networks and strong numerical samplers. Yet the two components are not fully matched. During sampling, the denoiser predicts for the next step based on flawed input that differ much from the real data. This train-test gap, or exposure bias, caps image generation quality (i.e., blurry, lacking in details or awkward in shape). To mitigate the problem, we propose PEEK, a novel training-free deterministic sampler plugin. At each eligible denoising step, PEEK first performs a provisional one-step update, evaluates the denoiser once at the look-ahead state, and interpolates it with the one-step-later prediction to form the next noisy state. By taking the additional step, we obtain a generally more-image-like prediction from the higher-signal noisy state, mitigating the exposure bias. To optimize the interpolation positions, PEEK searches for an interpolation schedule search made efficient by the prefix and baseline noise-schedule-rebasing caches. On 8 pixel-space and latent-space diffusion models with unconditional, label-guided, and text-guided settings across various output resolutions, PEEK consistently improves over NFE-matched baselines, with FID improvements up to 49.3\% at minimal extra computation cost (e.g., the one-time-per-model calibration takes at most 2.2 minutes across our experiments.). Project website is at this https URL.
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
arXiv:2611.05407 [cs.CV]
(or arXiv:2611.05407v1 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2611.05407 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 299 cuvinte