Research map / Image generation / Fast diffusion sampling
CL-DiffPhyCon: Closed-loop Diffusion Control of Complex Physical Systems
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 CL-DiffPhyCon, an asynchronous-denoising diffusion control method that re-plans each physical step from real-time system feedback, giving closed-loop control with lower sampling cost, and further accelerated by DDIM.
Problem
Closed-loop control of complex physical systems with diffusion-based generative controllers is costly and inefficient: synchronous full denoising per step is expensive, while infrequent replanning uses outdated states.
Approach
Decouple denoising across the model horizon so that different physical time steps have different noise levels (asynchronous diffusion); each step denoises only one new latent level, so the fresh system state conditions the next control signal without a full sampling pass.
Key results
- On 1D Burgers' control, CL-DiffPhyCon reduces the control objective to 0.0096 in noise-free setting vs 0.0210 for best baseline DiffPhyCon-1 (54.3% reduction), at 4.55 s average time.
- With DDIM (30 steps), CL-DiffPhyCon achieves about 5x speedup over RDM-level sampling time on 2D fluid control (26.01 s vs 238.43 s for RDM), with comparable control performance.
- On 2D fluid control (large domain, fixed map), CL-DiffPhyCon gets 0.3371 vs 0.4074 for the best baseline RDM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use asynchronous closed-loop denoising instead of DiffPhyCon-1 | DiffPhyCon · 1D Burgers' control, all six settings | control objective: 0.0210 -> 0.0096; 0.0214 -> 0.0110; 0.0222 -> 0.0095; 0.0232 -> 0.0127; 0.0330 -> 0.0291; 0.0332 -> 0.0295 | Helped |
| use asynchronous closed-loop denoising instead of the strongest baseline | DiffPhyCon · 2D fluid control, large domain and boundary control under fixed and random maps | control objective: 0.4074 -> 0.3371; 0.3754 -> 0.3485; 0.6498 -> 0.6169; 0.7087 -> 0.7003 | Helped |
| use asynchronous closed-loop denoising rather than RDM | Diffusion control · 2D fluid control (average across settings) | inference time: 238.43 -> 144.04 | Helped |
| use asynchronous closed-loop denoising rather than PID | Control method · 1D Burgers' control (average across settings) | inference time: 0.7717 -> 4.5474 | Hurt |
| add DDIM fast sampling (ablation) | CL-DiffPhyCon · 1D Burgers' control (average across settings) | inference time: 4.5474 -> 0.8257 | Helped |
| add DDIM fast sampling (ablation) | CL-DiffPhyCon · 1D Burgers' noise-free | control objective: 0.0096 -> 0.0112 | Hurt |
| add DDIM fast sampling (ablation) | CL-DiffPhyCon · 2D fluid control (average across settings) | inference time: 144.04 -> 26.01 | Helped |
| add DDIM fast sampling (ablation) | CL-DiffPhyCon · 2D fluid large domain, fixed map | control objective: 0.3371 -> 0.4100 | Hurt |
Limitations
- Trained offline without interaction with system dynamics
- No formal bound on optimization performance under guidance sampling
- Evaluated only on two physical control tasks
Experimental setup
Datasets: offline simulated trajectories of 1D Burgers' equation offline simulated trajectories of 2D incompressible fluid
Benchmarks: 1D Burgers' equation control 2D incompressible fluid control
Metrics: control objective (lower is better) average inference time (s)
Baselines: PID BC BPPO DiffPhyCon-1 DiffPhyCon-5 RDM
Builds on: DiffPhyCon DDPM DDIM RDM
Closest papers by meaning
- DiffPhyCon: A Generative Approach to Control Complex Physical Systems (NeurIPS 2024 · similarity 0.90)
- Safe and Stable Control via Lyapunov-Guided Diffusion Models (NeurIPS 2025 · similarity 0.85)
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.84)
- Adaptive Stochastic Coefficients for Accelerating Diffusion Sampling (NeurIPS 2025 · similarity 0.84)
- Stochastic Control for Fine-tuning Diffusion Models: Optimality, Regularity, and Convergence (ICML 2025 · similarity 0.83)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.83)
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.