Research map / Image generation / Fast diffusion sampling
DiffuserLite: Towards Real-time Diffusion Planning
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 DiffuserLite, a coarse-to-fine plan refinement process (PRP) for diffusion planning that generates short, jumpy plans level by level, achieving ~122 Hz decision frequency with competitive offline RL performance.
Problem
Diffusion planning methods make decisions at very low frequency (about 1 Hz) because iterative sampling of long-horizon trajectories with heavy networks is expensive, which blocks real-time control.
Approach
Plan-refinement process: generate a coarse plan of key points at equal intervals, then recursively refine only the first interval at each level with a small DiT-based diffusion model (or rectified flow), so redundant distant details are never modeled; the critic conditions each level via cumulative reward or value.
Key results
- Diffuser-Lite-R2 reaches an average decision frequency of 122.44 Hz, versus 1.34 Hz for Diffuser and 0.41 Hz for Decision Diffuser on D4RL.
- Diffuser-Lite-D averages 86.0 on D4RL medium-expert/medium/medium-replay vs 81.8 for DD and 82.6 for HDMI, and 74.0 on Kitchen mixed/partial vs 61.0 for DD.
- Integrating Diffuser-Lite into AlignDiff raises frequency from 6.9 Hz to 45.5 Hz with a 3.2% drop in MAE area.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply DiffuserLite planning to D4RL MuJoCo | DiffuserLite-D · D4RL MuJoCo average | normalized score: 82.6 -> 86.0 | Helped |
| apply DiffuserLite planning to D4RL Kitchen | DiffuserLite-D · D4RL Kitchen average | normalized score: 61.0 -> 74.0 | Helped |
| apply DiffuserLite planning to D4RL Antmaze | DiffuserLite-R1 · D4RL Antmaze average | success rate: 53.0 -> 82.5 | Helped |
| use rectified flow with reflow instead of Diffuser | D4RL runtime per action · D4RL | runtime (second): 0.749 -> 0.008 | Helped |
| apply DiffuserLite as an AlignDiff plugin | AlignDiff · AlignDiff | MAE Area: 0.621±0.023 -> 0.601±0.018; (3.2%↓) | Hurt |
| apply DiffuserLite as an AlignDiff plugin | AlignDiff · AlignDiff | Frequency (Hz): 6.9 -> 45.5; (560%↑) | Helped |
| use only the last planning level instead of all PRP levels (ablation) | DiffuserLite-D · Hopper-me | normalized score: 111.6±0.2 -> 27.8±10.2 | Hurt |
| remove progressive refinement planning (ablation) | DiffuserLite-D · Hopper-me | normalized score: 111.6±0.2 -> 96.6±1.0 | Hurt |
| use DiffuserLite network parameters and sampling steps in DD (ablation) | DD · Hopper-me | normalized score: 111.6±0.2 -> 67.9±24.7 | Hurt |
| increase model size from 2.7M to 4.22M parameters (ablation) | DiffuserLite-D · HalfCheetah-me | normalized score: 88.5±0.4 -> 90.6±1.1 | Helped |
| reduce sampling steps from 3 to 1 (ablation) | DiffuserLite-D · HalfCheetah-me | normalized score: 88.5±0.4 -> 74.4±15.7 | Hurt |
| increase sampling steps from 3 to 10 (ablation) | DiffuserLite-D · HalfCheetah-me | normalized score: 88.5±0.4 -> 89.2±1.3 | Helped |
Surprising findings
- A one-level model with only the last (short) horizon performs far worse (16.4 vs 103.1 on Hopper), showing the long-horizon coarse plan matters.
- Reducing DiT size from 4.22M to 0.68M parameters changes runtime only slightly, suggesting runtime is dominated by network forward-pass count rather than size.
Limitations
- Relies on classifier-free guidance, which requires tuning the target condition and becomes cumbersome across multiple levels.
- Temporal jump and planning-level choices are hand-designed; no automatic jump adjustment.
- HDMI was reimplemented by the authors since its code is not public, so its runtime and results are approximate.
Experimental setup
Datasets: D4RL Gym-MuJoCo D4RL FrankaKitchen D4RL Antmaze Robomimic PH FinRL
Benchmarks: D4RL Robomimic FinRL
Metrics: normalized return / success rate decision-making frequency (Hz) runtime per action (s) MAE area
Baselines: BC CQL IQL Diffuser Decision Diffuser HDMI BCQ IRIS MB-PPO AlignDiff GC
Builds on: Diffuser Decision Diffuser DiT IQL rectified flow
Closest papers by meaning
- Resisting Stochastic Risks in Diffusion Planners with the Trajectory Aggregation Tree (ICML 2024 · similarity 0.89)
- Simple Hierarchical Planning with Diffusion (ICLR 2024 · similarity 0.89)
- DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising (NeurIPS 2026 · similarity 0.88)
- What Makes a Good Diffusion Planner for Decision Making? (ICLR 2025 · similarity 0.87)
- Model-based Diffusion for Trajectory Optimization (NeurIPS 2024 · similarity 0.87)
- Monte Carlo Tree Diffusion for System 2 Planning (ICML 2025 · similarity 0.87)
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.