English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Diffusion models achieve great success in generating diverse andhigh-fidelity images. The performance improvements come with low generationspeed per image, which hinders the application diffusion models in real-timescenarios. While some certain predictions benefit from the full computation ofthe model in each sample iteration, not every iteration requires the sameamount of computation, potentially leading to computation waste. In this work,we propose DeeDiff, an early exiting framework that adaptively allocatescomputation resources in each sampling step to improve the generationefficiency of diffusion models. Specifically, we introduce a timestep-awareuncertainty estimation module (UEM) for diffusion models which is attached toeach intermediate layer to estimate the prediction uncertainty of each layer.The uncertainty is regarded as the signal to decide if the inferenceterminates. Moreover, we propose uncertainty-aware layer-wise loss to fill theperformance gap between full models and early-exited models. With such lossstrategy, our model is able to obtain comparable results as full-layer models.Extensive experiments of class-conditional, unconditional, and text-guidedgeneration on several datasets show that our method achieves state-of-the-artperformance and efficiency trade-off compared with existing early exitingmethods on diffusion models. More importantly, our method even brings extrabenefits to baseline models and obtains better performance on CIFAR-10 andCeleb-A datasets. Full code and model are released for reproduction.

DeeDiff: Dynamic Uncertainty-Aware Early Exiting for Accelerating Diffusion Model Generation