Research map / Image generation / Fast diffusion sampling

Entropic Time Schedulers for Generative Diffusion Models

NeurIPS 2025 · Image generation › Fast diffusion sampling · approach: Training-free diffusion acceleration · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes entropic time (conditional entropy of x0 given xt, estimated from the denoising loss) and a rescaled variant as data-dependent time reparameterizations for diffusion sampling schedules, showing FID/FD-DINOv2 gains on pretrained EDM/EDM2 models at low NFE.

Problem

Choosing the noise schedule (time reparameterization) for diffusion models strongly affects sample quality and efficiency, especially at few sampling steps, but standard schedules are heuristic and data-independent.

Approach

Define time as the (rescaled) conditional entropy H[x0|xt], which yields a constant entropy rate so each sampling point contributes equally information; show it is invariant to the original time parameterization and compute it exactly from the denoising score-matching/EDM loss without extra training.

Key results

• On ImageNet-512 with EDM2-XXL and stochastic DDIM at NFE=16, rescaled entropy schedule reduces FID from 30.39 (EDM) to 13.38 and FD-DINOv2 from 218.10 to 108.16.
• On FFHQ 64x64 with deterministic DDIM at NFE=16, FID drops from 11.13 (EDM) to 8.10 (rescaled entropy).
• Entropic (non-rescaled) time produced blurry, unrecognizable images on image data and was excluded from the main comparison; gains are largest in the few-NFE regime and mostly for first-order solvers.

Measured effects reported in the paper

Surprising findings

• The non-rescaled entropic time produced unrecognizable images on image data, so only the rescaled variant was used.
• Second-order solvers performed worse with entropic schedules than with the EDM schedule.

Limitations

• Benefits observed mainly for first-order solvers; second-order solvers gave worse results than the EDM schedule.
• Entropy estimated from the trained model's error ignores the learned-vs-true score discrepancy, making it an upper bound approximation.
• Non-rescaled entropic time gave blurry images on image data; theory for discrete diffusion conjectured but not proven.

Experimental setup

Datasets: CIFAR-10 FFHQ ImageNet-64 ImageNet-512 mixture of Gaussians (1D) discrete 15-point mixture (1D)

Benchmarks: CIFAR-10 FFHQ 64x64 ImageNet-64 ImageNet-512

Metrics: FID FD-DINOv2 KL divergence

Baselines: EDM schedule standard (uniform) time schedule entropic time deterministic DDIM stochastic DDIM

Builds on: EDM DDIM Karras et al. optimal schedule (theorem 3.1) time change of SDEs I-MMSE identity

Closest papers by meaning

• Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training (CVPR 2025 · similarity 0.90)
• Denoising Task Difficulty-based Curriculum for Training Diffusion Models (ICLR 2025 · similarity 0.89)
• Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.88)
• Align Your Steps: Optimizing Sampling Schedules in Diffusion Models (ICML 2024 · similarity 0.88)
• EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.88)
• ART for Diffusion Sampling: A Reinforcement Learning Approach to Timestep Scheduling (NeurIPS 2026 · similarity 0.88)