Research map / Image generation / Fast diffusion sampling
MosaicDiff: Training-free Structural Pruning for Diffusion Model Acceleration Reflecting Pretraining Dynamics
Bowei Guo, Shengkun Tang, Cong Zeng, Zhiqiang Shen
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 SNR-aware calibration data (ablation) | MosaicDiff · DiT-XL/2 20 steps | FID: 3.96 -> 3.22 | Improved |
| add SNR-aware calibration data (ablation) | MosaicDiff · DiT-XL/2 20 steps | IS: 227.3 -> 266.1 | Improved |
| add score-based three-stage division instead of no division (ablation) | MosaicDiff · DiT-XL/2 20 steps, sparsity 0.3 | FID: 40.89 -> 3.20 | Improved |
| use proposed stage division instead of uniform division (ablation) | MosaicDiff · DiT-XL/2 20 steps | FID: 33.21 -> 3.20 | Improved |
| choose threshold M=0.55 instead of 0.45 (ablation) | MosaicDiff · DiT-XL/2 20 steps | FID: 3.36 -> 3.20 | Improved |
| choose threshold M=0.55 instead of 0.70 (ablation) | MosaicDiff · DiT-XL/2 20 steps | FID: 3.71 -> 3.20 | Improved |
| increase sparsity in Stage 2 (0.04 to 0.15) (ablation) | MosaicDiff · DiT-XL/2 20 steps | FID: 3.22 -> 10.20 | Hurt |
| increase sparsity in Stage 3 (0.1 to 0.4) without SNR refinement (ablation) | MosaicDiff · DiT-XL/2 20 steps | FID: 3.22 -> 4.18 | Hurt |
| increase sparsity in Stage 1 (0.6 to 0.9) (ablation) | MosaicDiff · DiT-XL/2 20 steps | FID: 3.22 -> 3.53 | No clear effect |
| raise total sparsity from 0.30 to 0.50 (ablation) | MosaicDiff · DiT-XL/2 20 steps | FID: 3.20 -> 5.27 | Hurt |
| raise total sparsity from 0.10 to 0.30 (ablation) | MosaicDiff · SDXL COCO | FID-5K: 23.18 -> 28.37 | Hurt |
| add MosaicDiff pruning to DeepCache-N=3 | DeepCache · SDXL | FID-5K: 24.61 -> 23.88 | Improved |
Later papers that compared with this method
1 later accepted paper reports MosaicDiff as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- BiGain: Unified Token Compression for Joint Generation and Classification (CVPR 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation (NeurIPS 2025 · similarity 0.88)
- PFDiff: Training-Free Acceleration of Diffusion Models Combining Past and Future Scores (ICLR 2025 · similarity 0.88)
- Morse: Dual-Sampling for Lossless Acceleration of Diffusion Models (ICML 2025 · similarity 0.88)
- DiP-GO: A Diffusion Pruner via Few-step Gradient Optimization (NeurIPS 2024 · similarity 0.88)
- T-Stitch: Accelerating Sampling in Pre-Trained Diffusion Models with Trajectory Stitching (ICLR 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.
