Research map / Image generation / Fast diffusion sampling
ParaLin: Accelerating Parallel Diffusion Integrator via Intrinsic Partially Linear Structure
Jianrong Lu, Zhiyu Zhu, Hui LIU, Junhui Hou
Read the paperCheck your idea against 137,217 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.
Accelerates diffusion sampling by reformulating sequential integral solvers as a parallel process exploiting an intrinsic partially linear structure (ParaLin).
Problem
Sequential inference of diffusion probabilistic models is slow because each denoising step depends on the previous one.
Approach
Views diffusion sampling as a dynamical system; shows the sequential integral solver can be approximated by a linear solver, enabling parallel sampling that exploits a partially linear structure.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use 16-step sequential UniPC/DPMSolver++ instead of 30-step Euler | FLUX.1-dev · GenEval | GenEval: 0.66 -> 0.50-0.52 | Hurt |
| use 16-step sequential UniPC/DPMSolver++ instead of 30-step Euler | FLUX.1-dev · DPG-Bench | DPG-Bench: 83.97 -> 75.75-77.46 | Hurt |
| use 16-step sequential UniPC/DPMSolver++ instead of 30-step Euler | FLUX.1-dev · LongTextBench | LongTextBench: 0.27 -> 0.08-0.09 | Hurt |
| keep 30 discrete steps and parallelize solver (ParaLin++) | FLUX.1-dev · GenEval/DPG/LongTextBench | quality profile: 0.64 vs 0.66 GenEval, 84.24 vs 83.97 DPG, 0.26 vs 0.27 LTB | No clear effect |
| keep 30 discrete steps and parallelize solver (ParaLin++) | Ming-Flash-Omni 2.0 · GenEval | GenEval: 0.88 -> 0.87 | No clear effect |
| increase number of sequential steps for few-step solvers (16 to 30) (ablation) | UniPC on FLUX.1-dev · GenEval | GenEval: 0.52 -> 0.60 | Improved |
| vary CFG from 0 to 14 for 20-step few-step solvers (ablation) | UniPC and DPMSolver++ on FLUX.1-dev · DPG-Bench | DPG-Bench: 68.85-79.67 vs Euler 82.25-83.97 | Hurt |
| add parallel solver ParaLin++ to DDIM | Stable Diffusion-v2, 50 steps · wall-clock speedup | speedup: 1.0x -> 4.8x | Improved |
| use full nonlinear parallel system instead of partial nonlinearity (ablation) | ParaLin++ on 100-step DDIM · ablation | NFE and speedup: partial nonlinearity removes 93.3% of NFEs | Hurt |
| add adaptive stopping criterion (ablation) | ParaLin++ on 100-step DDIM · ablation | iterations and NFE | Improved |
| add successive over-relaxation (SOR) (ablation) | ParaLin++ on 100-step DDIM · ablation | speedup: 15 iterations, 117 NFEs, 6.5x speedup | Improved |
| run on a single GPU | ParaDiGMS vs ParaLin++ · Stable Diffusion-v2 DDIM-100 | speedup: ParaDiGMS 0.8x; ParaLin++ 4.2x | Mixed |
Experimental setup
Builds on: DDIM Picard iteration ParaDiGMS
Closest papers by meaning
- Accelerating Diffusion Models with Parallel Sampling: Inference at Sub-Linear Time Complexity (NeurIPS 2024 · similarity 0.91)
- Self-Refining Diffusion Samplers: Enabling Parallelization via Parareal Iterations (NeurIPS 2024 · similarity 0.91)
- A Unified Sampling Framework for Solver Searching of Diffusion Probabilistic Models (ICLR 2024 · similarity 0.89)
- PFDiff: Training-Free Acceleration of Diffusion Models Combining Past and Future Scores (ICLR 2025 · similarity 0.89)
- Accelerating Parallel Diffusion Model Serving with Residual Compression (NeurIPS 2025 · similarity 0.88)
- Accelerating Convergence of Score-Based Diffusion Models, Provably (ICML 2024 · similarity 0.88)
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.
