AI Paper Atlas

Research map / Image generation / Fast diffusion sampling

Data-to-Energy Stochastic Dynamics

Kirill Tamogashev, Nikolay Malkin

ICLR 2026 · Image generation › Fast diffusion sampling · approach: Diffusion theory and learned samplers

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
learn diffusion variance in addition to drift (ablation)data-to-data IPF (log-likelihood) · Two Moons<->GMM (K=20)W2^2: 0.031 -> 0.023Improved
learn diffusion variance in addition to drift (ablation)data-to-data IPF (log-likelihood) · Gauss<->Two Moons (K=20)W2^2: 0.033 -> 0.022Improved
learn diffusion variance (ablation)data-to-data IPF · Gauss<->GMM (K=20)W2^2: 0.037 -> 0.042Hurt
learn diffusion variance (ablation)data-to-data IPF · Gauss<->GMM (K=5)W2^2: 0.066 -> 0.026Improved
learn diffusion variance (ablation)data-to-data IPF · Gauss<->GMM (K=5)Path KL: 0.988 -> 4.158Hurt
learn diffusion variance with large step count (ablation)data-to-data IPF · Gauss<->GMM (K=40)W2^2: 0.023 -> 0.028No clear effect
add replay buffer to on-policy training (ablation)data-to-energy IPF · CIFAR-10 SN-GAN dogsPath KL: 1506.407 -> 622.895Improved
add Langevin updates on buffer (ablation)data-to-energy IPF with buffer · CIFAR-10 SN-GAN dogsPath KL: 622.895 -> 383.514Improved
reuse backward trajectories in loss (ablation)data-to-energy IPF with buffer and Langevin · CIFAR-10 SN-GAN dogsPath KL: 383.514 -> 206.094Improved
reuse backward trajectories in loss (ablation)data-to-energy IPF with buffer and Langevin · CIFAR-10 SN-GAN dogsmean log-reward: -0.286 -> -0.657Hurt
use smaller off-policy ratio (ablation)data-to-energy IPF · CIFAR-10 SN-GAN dogsPath KL: 206.094 -> 668.255Hurt
anneal off-policy ratio (ablation)data-to-energy IPF · CIFAR-10 SN-GAN dogsmean log-reward: -0.657 -> -0.149Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy