Research map / Image generation / Fast diffusion sampling

AsyncDiff: Parallelizing Diffusion Models by Asynchronous Denoising

NeurIPS 2024 · Image generation › Fast diffusion sampling · approach: Training-free diffusion acceleration · 7 citations (OpenAlex)

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 AsyncDiff, a training-free model-parallel scheme that splits the denoising U-Net into components on separate GPUs and uses previous-step hidden states to break sequential dependencies, plus stride denoising to cut communication.

Problem

Multi-step sequential denoising in diffusion models causes high cumulative latency and prevents parallel computation across devices; prior parallel methods (patch parallelism, Picard iterations) have low utilization, high memory, or quality loss.

Approach

Partition the noise prediction network into sequential components on different devices and feed each component the previous step's output from the preceding component, so the components run asynchronously on different timesteps; stride denoising further skips redundant computation and reduces broadcast frequency.

Key results

• On SD 2.1 with 4 A5000 GPUs, AsyncDiff (N=3, S=2) reaches 4.0x speedup (1.35s vs 5.51s) with CLIP Score 31.22 vs 31.60 original.
• On SDXL, N=3 S=2 on 4 devices gives 3.8x speedup (3.65s vs 13.81s) with CLIP 31.40 vs 32.33.
• At matched 2.3x and 2.7x speedups, AsyncDiff shows better FID/LPIPS than DistriFusion while using less memory (about 5.5GB vs 6.5-7.3GB).

Measured effects reported in the paper

Surprising findings

• Speedups can exceed the number of devices (e.g., 4.3x on 4 devices for SD 2.1 with 100 steps) because stride denoising skips redundant computation.

Limitations

• Requires frequent inter-device communication, so poor interconnects degrade performance.
• Depends on quality of the pre-trained base model; cannot exceed its baseline quality.
• Higher speedups reduce pixel-level consistency with original outputs (LPIPS rises to ~0.38).

Experimental setup

Datasets: MS-COCO 2017 validation

Benchmarks: MS-COCO 2017 validation

Metrics: CLIP Score FID LPIPS latency speedup MACs DISTS MUSIQ CLIP-IQA NIQE

Baselines: DistriFusion Faster Diffusion ParaDiGMS Original model (sequential inference)

Builds on: DDIM DistriFusion Faster Diffusion ParaDiGMS

Closest papers by meaning

• Communication-Efficient Diffusion Denoising Parallelization via Reuse-then-Predict Mechanism (NeurIPS 2025 · similarity 0.92)
• DistriFusion: Distributed Parallel Inference for High-Resolution Diffusion Models (CVPR 2024 · similarity 0.91)
• Accelerating Diffusion Models with Parallel Sampling: Inference at Sub-Linear Time Complexity (NeurIPS 2024 · similarity 0.90)
• Accelerating Parallel Diffusion Model Serving with Residual Compression (NeurIPS 2025 · similarity 0.90)
• Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.89)
• EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.89)