Research map / Image generation / Fast diffusion sampling
Structural Pruning for Diffusion Models
Gongfan Fang, Xinyin Ma, Xinchao Wang
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 |
|---|---|---|---|
| Use Diff-Pruning with T=0.05 instead of the strongest listed pruning baseline | CIFAR-10 diffusion model · CIFAR-10 | FID: 5.48 -> 5.29 | Improved |
| Use Diff-Pruning with T=0.00 instead of the strongest listed pruning baseline | CelebA-HQ diffusion model · CelebA-HQ | FID: 6.64 -> 6.24 | Improved |
| Prune the pretrained LSUN-Church model with Diff-Pruning | LSUN-Church DDPM · LSUN-Church | FID: 10.6 -> 13.9 | Hurt |
| Prune the pretrained LSUN-Bedroom model with Diff-Pruning | LSUN-Bedroom DDPM · LSUN-Bedroom | FID: 6.9 -> 18.6 | Hurt |
| Use Diff-Pruning instead of Taylor pruning | ImageNet-1K conditional LDM · ImageNet-1K 256×256 | FID: 11.18 -> 9.16 | Improved |
| Add pruning to the pretrained CIFAR-10 model at 16% pruning (ablation) | CIFAR-10 diffusion model · CIFAR-10 | FID: 4.19 -> 4.62 | Hurt |
| Increase pruning ratio from 44% to 70% (ablation) | CIFAR-10 Diff-Pruning · CIFAR-10 | FID: 5.29 -> 9.33 | Hurt |
| Prune uninformative timesteps using threshold T=0.05 rather than using all timesteps (ablation) | CIFAR-10 Diff-Pruning · CIFAR-10 | FID: 5.49 -> 5.29 | Improved |
| Increase the number of Taylor timesteps beyond about 250 (ablation) | CIFAR-10 pruning importance estimation · CIFAR-10 | SSIM | Hurt |
| Use more Taylor timesteps for importance estimation (ablation) | CelebA-HQ pruning importance estimation · CelebA-HQ | SSIM | Improved |
| Train the compact model from scratch instead of pruning the pretrained model | Compact diffusion model · CIFAR-10 and CelebA-HQ | FID | Hurt |
Closest papers by meaning
- ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation (NeurIPS 2025 · similarity 0.90)
- DisCo-Diff: Enhancing Continuous Diffusion Models with Discrete Latents (ICML 2024 · similarity 0.90)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.90)
- Efficient Fine-Tuning and Concept Suppression for Pruned Diffusion Models (CVPR 2025 · similarity 0.89)
- DiP-GO: A Diffusion Pruner via Few-step Gradient Optimization (NeurIPS 2024 · similarity 0.89)
- Sparsely Supervised Diffusion (NeurIPS 2026 · similarity 0.89)
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.
