Research map / Image generation / Fast diffusion sampling
CleanDiffuser: An Easy-to-use Modularized Library for Diffusion Models in Decision Making
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.
Introduces CleanDiffuser, a modular open-source library for diffusion-model-based decision making that decouples diffusion backbones, network architectures, and guided sampling, and benchmarks 9 DM algorithms across 37 RL/IL tasks.
Problem
Diffusion-model-based decision-making algorithms lack an easy-to-use, modular open-source library; existing codebases are algorithm-specific and tightly coupled, hindering customization and reproducible comparison.
Approach
Decompose DM-based decision-making into decoupled sub-modules (diffusion SDE/ODE and solvers, network architectures, guided sampling) plus environment wrappers and dataloader, so algorithms are assembled from building blocks in about 10 lines of code; a trained model can switch solvers and step counts at inference without retraining.
Key results
- Across D4RL offline RL tasks, DQL achieves the highest average score of 89.0 among the diffusion-based methods evaluated.
- On Robomimic/PushT IL benchmarks, DiffusionPolicy with Chi_UNet1d reaches an average of 0.90/0.96 (last/max) on low-dim tasks, the best among compared methods.
- Five sampling steps are adequate for most tasks, and sampling degradation (performance dropping with more steps) occurs in some tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase the EMA rate (ablation) | IDQL, DD, DiffusionBC, and DiffusionPolicy · training across evaluated tasks | performance and training stability | Helped |
| increase diffusion sampling steps (ablation) | IDQL and DD · medium-expert MuJoCo and Kitchen tasks | performance | Hurt |
| use SDE solvers instead of ODE solvers (ablation) | diffusion policies · evaluated decision-making tasks | performance | Helped |
| use a higher-order ODE solver rather than a first-order solver (ablation) | diffusion-based decision-making algorithms · evaluated tasks | performance | No clear effect |
| increase model size from 4M to 15M (ablation) | DD · Kitchen-m | normalized score: 56.5±5.8 -> 80.5±4.1 | Helped |
| increase model size from 15M to 60M (ablation) | DD · Kitchen-m | normalized score: 80.5±4.1 -> 27.7±2.1 | Hurt |
| increase model size from 4M to 15M (ablation) | DD · Antmaze-mp | success rate: 8.0±4.3 -> 26.0±5.9 | Helped |
| increase model size across tested sizes (ablation) | IDQL · HalfCheetah-m | normalized score: 51.5±0.1 -> 51.5±0.1 -> 51.7±0.1 | No clear effect |
| use DiffusionPolicy rather than DiffusionBC | diffusion imitation learning · evaluated offline IL tasks | performance | Helped |
| use Chi_UNet1d rather than other tested architectures (ablation) | DiffusionPolicy · evaluated offline IL tasks | performance | Helped |
| increase model size for Transformer-based diffusion models (ablation) | DiT1d · imitation learning tasks | performance | Mixed |
Surprising findings
- Increasing sampling steps can degrade performance in some tasks (sampling degradation).
- Diffusion planners show no superiority over diffusion policies and perform poorly on Antmaze.
Limitations
- Diffusion policies have slower inference than RNN or transformer IL baselines due to multiple denoising forwards.
- Sampling degradation and training degradation remain unexplained open issues.
- Larger models do not consistently improve performance.
Experimental setup
Datasets: D4RL PushT Relay Kitchen Robomimic
Benchmarks: D4RL PushT Relay Kitchen Robomimic
Metrics: normalized score success rate target area coverage
Baselines: BC LSTM-GMM ACT Diffuser Decision Diffuser DQL EDP IDQL SynthER AdaptDiffuser
Builds on: DDPM DDIM DPM-Solver EDM Diffuser Diffusion Policy
Closest papers by meaning
- Align Your Steps: Optimizing Sampling Schedules in Diffusion Models (ICML 2024 · similarity 0.87)
- Target Concrete Score Matching: A Holistic Framework for Discrete Diffusion (ICML 2025 · similarity 0.86)
- TFG: Unified Training-Free Guidance for Diffusion Models (NeurIPS 2024 · similarity 0.86)
- Diffusion Spectral Representation for Reinforcement Learning (NeurIPS 2024 · similarity 0.86)
- Habitizing Diffusion Planning for Efficient and Effective Decision Making (ICML 2025 · similarity 0.86)
- What Makes a Good Diffusion Planner for Decision Making? (ICLR 2025 · similarity 0.85)
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.