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

Due to the ease of training, ability to scale, and high sample quality,diffusion models (DMs) have become the preferred option for generativemodeling, with numerous pre-trained models available for a wide variety ofdatasets. Containing intricate information about data distributions,pre-trained DMs are valuable assets for downstream applications. In this work,we consider learning from pre-trained DMs and transferring their knowledge toother generative models in a data-free fashion. Specifically, we propose ageneral framework called Diff-Instruct to instruct the training of arbitrarygenerative models as long as the generated samples are differentiable withrespect to the model parameters. Our proposed Diff-Instruct is built on arigorous mathematical foundation where the instruction process directlycorresponds to minimizing a novel divergence we call Integral Kullback-Leibler(IKL) divergence. IKL is tailored for DMs by calculating the integral of the KLdivergence along a diffusion process, which we show to be more robust incomparing distributions with misaligned supports. We also reveal non-trivialconnections of our method to existing works such as DreamFusion, and generativeadversarial training. To demonstrate the effectiveness and universality ofDiff-Instruct, we consider two scenarios: distilling pre-trained diffusionmodels and refining existing GAN models. The experiments on distillingpre-trained diffusion models show that Diff-Instruct results instate-of-the-art single-step diffusion-based models. The experiments onrefining GAN models show that the Diff-Instruct can consistently improve thepre-trained generators of GAN models across various settings.

Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models