Research map / Image generation / Fast diffusion sampling
Communication-Efficient Diffusion Denoising Parallelization via Reuse-then-Predict Mechanism
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 ParaStep, a step-wise parallel diffusion sampler using a reuse-then-predict mechanism that exchanges only noise and sample tensors per denoising cycle, plus a single-device BatchStep variant for audio models.
Problem
Diffusion inference latency from sequential denoising; existing multi-device parallelization methods incur high layer-wise or stage-wise communication, limiting use on low-bandwidth commodity hardware.
Approach
Reuse the previous step's noise to form an approximate noisy sample for the current step, then run a fresh noise prediction on that sample; distribute adjacent-step predictions across GPUs so only noise/sample tensors are communicated once per step cycle.
Key results
- ParaStep achieves 3.88x speedup on SVD at parallelism degree 8 and 2.43x on CogVideoX-2b (per abstract), with 2.49x on SVD at p=4 in the main table.
- On SD3 with p=2, ParaStep reaches 1.68x speedup with FID 5.01, versus 7.11 for AsyncDiff and 6.20 for xDiT-Pipe.
- On AudioLDM2-large, ParaStep reaches 6.56x speedup at parallelism degree 8 with only a 0.02 FAD increase.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ParaStep instead of AsyncDiff at p=4 | SVD · VBench (14 frames, 1024×576) | speedup: 1.80 -> 2.49 | Helped |
| Use ParaStep instead of xDiT-Ring at p=4 | CogVideoX-2b · VBench (45 frames, 512×720) | speedup: 0.98 -> 1.93 | Helped |
| Use ParaStep instead of xDiT-Ring at p=4 | Latte · VBench (16 frames, 512×512) | VBench: 73.81 -> 73.76 | Hurt |
| Use ParaStep instead of xDiT-Ring at p=4 | HunyuanVideo · VBench (5 frames, 180×180) | speedup: 1.07 -> 2.67 | Helped |
| Increase parallelism degree to 8 (ablation) | ParaStep on AudioLDM2-large · AudioCaps | speedup: 6.56× speedup | Helped |
| Increase parallelism degree to 8 (ablation) | ParaStep on SVD · VBench | speedup: 3.88× speedup | Helped |
| Use 30% warm-up ratio rather than fewer warm-up steps (ablation) | ParaStep on CogVideoX-2b · VBench | generation quality | Helped |
| Increase warm-up ratio from 30% to 100% (ablation) | ParaStep on CogVideoX-2b · VBench | speedup | Hurt |
| Use BatchStep instead of ParaStep on one device | AudioLDM2-large · AudioCaps | generation latency | Helped |
| Replace direct noise reuse with reuse-then-predict | Diffusion denoising · Adjacent denoising steps | deviation from original model | Helped |
| Integrate TeaCache schedule into ParaStep (ablation) | ParaStep · Diffusion generation | generation quality | Helped |
Surprising findings
- Direct noise reuse degrades quality noticeably, but a single fresh prediction after reuse largely recovers it.
- For video models, the warm-up phase is long (13-18 steps), which limits speedup for CogVideoX-2b and Latte.
Limitations
- Replicates the full noise predictor on every GPU, so per-GPU memory is not reduced.
- Requires extra compute resources to trade for latency reduction.
- Speedup limited for models with many warm-up steps (e.g., CogVideoX-2b, Latte).
Experimental setup
Datasets: MS-COCO 2017 VBench AudioCaps
Benchmarks: MS-COCO 2017 VBench AudioCaps
Metrics: speedup latency FID LPIPS PSNR SSIM VBench FAD
Baselines: AsyncDiff xDiT-Pipe xDiT-Ring DeepCache TeaCache
Builds on: DistriFusion PipeFusion AsyncDiff TeaCache DeepCache
Closest papers by meaning
- AsyncDiff: Parallelizing Diffusion Models by Asynchronous Denoising (NeurIPS 2024 · similarity 0.92)
- Accelerating Parallel Diffusion Model Serving with Residual Compression (NeurIPS 2025 · similarity 0.90)
- Accelerating Diffusion Models with Parallel Sampling: Inference at Sub-Linear Time Complexity (NeurIPS 2024 · similarity 0.90)
- TRIM: Scalable 3D Gaussian Diffusion Inference with Temporal and Spatial Trimming (NeurIPS 2025 · similarity 0.90)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.90)
- EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.90)
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.