ART for Diffusion Sampling: Continuous-Time Control and Actor-Critic Learning

2026-07-02 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Adaptive Reparameterized Time (ART) is a novel continuous-time control formulation designed to optimize timestep allocation in score-based diffusion sampling. Addressing the suboptimality of fixed uniform or hand-crafted schedules, ART learns a time change by controlling the sampling clock's speed, enabling adaptive timesteps in the original diffusion time. It employs a principled objective based on a leading-order Euler error surrogate. To solve this deterministic control problem, the authors introduce ART-RL, an auxiliary randomized formulation utilizing Gaussian policies, which reframes schedule learning as a continuous-time reinforcement learning problem. ART-RL is proven to be equivalent to ART at the optimizer level, with its optimal Gaussian policy recovering the optimal ART time-warping rate. Experiments demonstrate that ART-RL consistently improves sample quality over strong baseline schedules at matched budgets, integrates seamlessly into existing diffusion samplers, and exhibits broad generalization across various sampling budgets, datasets, solvers, pipelines, and representation spaces.

Key takeaway

For Machine Learning Engineers optimizing diffusion model inference, ART-RL offers a significant advancement. You can integrate this method into your existing diffusion samplers by only changing the timestep grid. This approach consistently improves sample quality at matched computational budgets. Furthermore, the learned schedules generalize broadly, reducing the need for retraining across different datasets, solvers, or budgets. Consider adopting ART-RL to enhance your diffusion model performance and efficiency.

Key insights

ART-RL optimizes diffusion sampling timesteps via continuous-time control and actor-critic learning, improving sample quality and generalization.

Principles

• Optimal timestep allocation improves diffusion sample quality.
• Continuous-time control can optimize discrete processes.
• RL can solve deterministic control problems via randomization.

Method

ART-RL formulates timestep allocation as a continuous-time RL problem, learning a time-warping rate via Gaussian policies and actor-critic updates, based on an Euler error surrogate.

In practice

• Plug ART-RL into existing diffusion samplers.
• Improve sample quality at matched computational budgets.
• Transfer learned schedules across diverse settings.

Topics

• Diffusion Sampling
• Timestep Allocation
• Continuous-Time Control
• Reinforcement Learning
• Actor-Critic Learning
• Score-based Diffusion Models

Best for: Research Scientist, Computer Vision Engineer, AI Scientist, Machine Learning Engineer

Related on AIssential

• Enhanced Low-Density Region Exploration in Classifier-Guided Diffusion Models Through Modified Reverse Diffusion Sampling — A novel sampling-time modification for classifier-guided diffusion models improves low-density region exploration without retraining.
• PTL-Diffusion: Manifold-Aware Diffusion with Periodic Terminal Laws — PTL-Diffusion improves manifold data generation by using periodic, structured terminal laws in its forward noising process.
• The Race to Production-Grade Diffusion LLMs with Stefano Ermon - #764 — Diffusion language models offer significantly faster and more efficient inference for text and code generation than autoregressive LLMs.
• Temporal Difference Learning for Diffusion Models — Temporal difference learning improves diffusion model consistency and sample quality, especially for few-step sampling.
• A unified perspective on fine-tuning and sampling with diffusion and flow models — A unified framework clarifies fine-tuning and sampling in diffusion/flow models via exponential tilting.
• A Survey on Diffusion Language Models — Diffusion Language Models (DLMs) achieve parallel text generation and bidirectional context through iterative denoising, rivaling autoregressive models in performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.