Research map / Image generation / Fast diffusion sampling
Exact simulation of diffusion first exit times: algorithm acceleration
Samuel Herrmann, Cristina Zucca
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-armed bandit choice of box parameter N (eps=0.5) | BoxExit exit-time simulation · Example 1 diffusion, drift 2+sin(x), [0,7], x=3 | computation time: 5.2% of the classical consumption time | Improved |
| add multi-armed bandit choice of N with small exploration (eps=0.1) | BoxExit exit-time simulation · Example 1 diffusion, drift 2+sin(x), [0,7], x=3 | computation time: 2.1% of the consumption time without acceleration | Improved |
| add purely uniform random choice of N (eps=1) (ablation) | BoxExit exit-time simulation · Example 1 diffusion, drift 2+sin(x), [0,7], x=3 | computation time: 9.7% of the classical consumption time | Improved |
| lower exploration rate eps from 1 to 0.1 (ablation) | BanditDiffExit · Example 1 diffusion, drift 2+sin(x), [0,7], x=3 | share of runs with optimal N=14: 5.1% -> 59.7% | Improved |
| use known optimal N=14 instead of bandit (ablation) | DiffExit · Example 1 diffusion, drift 2+sin(x), [0,7], x=3 | computation time: 0.15% of the classical time consumption | Improved |
| increase T from 0.5 to 3 (ablation) | BoxExit on OU Ex.1 · OU Ex.1 | computation time (ms): 4.195 -> 13.135 | Hurt |
| increase T from 0.5 to 3 (ablation) | BoxExit on OU Ex.2 · OU Ex.2 | computation time (ms): 46.471 -> 186.576 | Hurt |
| decrease T from 0.5 to 0.1 (ablation) | BoxExit on OU Ex.2 · OU Ex.2 | computation time (ms): 46.471 -> 68.636 | Hurt |
| vary T between 0.1 and 1 (ablation) | BoxExit on CIR · CIR model | computation time (ms): 12.206 vs 11.901 | No clear effect |
| add bandit N selection with small eps | BanditDiffExit on CIR · CIR model | computation time per simulation (ms): about 12 ms -> 0.2 ms | Improved |
| have origin inside the interval [a,b] (ablation) | BanditDiffExit on OU · OU Ex.2 vs Ex.1 | acceleration strength: N=14 optimal -> N=5 optimal, boxes 1 800 scale | Hurt |
Closest papers by meaning
- Accelerating Convergence of Score-Based Diffusion Models, Provably (ICML 2024 · similarity 0.81)
- Self-Refining Diffusion Samplers: Enabling Parallelization via Parareal Iterations (NeurIPS 2024 · similarity 0.81)
- On Accelerating Diffusion-Based Sampling Processes via Improved Integration Approximation (ICLR 2024 · similarity 0.81)
- Adaptive Stochastic Coefficients for Accelerating Diffusion Sampling (NeurIPS 2025 · similarity 0.81)
- A Unified Sampling Framework for Solver Searching of Diffusion Probabilistic Models (ICLR 2024 · similarity 0.81)
- A Score-Based Deterministic Diffusion Algorithm with Smooth Scores for General Distributions (AAAI 2024 · similarity 0.81)
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.
