Research map / Image generation / Fast diffusion sampling
Pluggable Pruning with Contiguous Layer Distillation for Diffusion Transformers
Jian Ma, Qirong Peng, Xujie Zhu, Peixing Xie, Chen Chen, Haonan Lu
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 linear probing layer selection (ablation) | PPCL · Qwen-Image 12B ablation | average score: 0.706 -> 0.761 | Improved |
| use first-order difference instead of CKA threshold for interval end (ablation) | PPCL · Qwen-Image 12B ablation | average score: 0.718 -> 0.761 | Improved |
| extend intervals by one layer (LP-b) (ablation) | PPCL · Qwen-Image 12B ablation | average score: 0.761 -> 0.726 | Hurt |
| replace sequential with non-sequential distillation (ablation) | PPCL · Qwen-Image 12B ablation | average score: 0.761 -> 0.848 | Improved |
| add text-stream width pruning (linear layers) (ablation) | PPCL · Qwen-Image ablation | average score: 0.848 -> 0.860 | Improved |
| replace FFN with linear layer (ablation) | PPCL · Qwen-Image ablation | average score: 0.860 -> 0.850 (1B fewer params) | Mixed |
| add full-parameter fine-tuning (ablation) | PPCL · Qwen-Image ablation | average score: 0.850 -> 0.870 | Improved |
| PPCL vs TinyFusion/HierarchicalPrune at 8B | PPCL · FLUX.1-dev | avg performance drop (%): 13.80 / 13.38 -> 4.03 | Improved |
| PPCL vs TinyFusion at 14B | PPCL · Qwen-Image | avg performance drop (%): 8.75 -> 0.42 | Improved |
| restore teacher layers in 10B model to get 12B/14B (plug-and-play) | PPCL · Qwen-Image | avg performance drop (%): 6.88 (10B) -> 3.03 (12B) -> 0.42 (14B) | Improved |
Closest papers by meaning
- TinyFusion: Diffusion Transformers Learned Shallow (CVPR 2025 · similarity 0.90)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.89)
- FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification (NeurIPS 2024 · similarity 0.89)
- Adaptive Pruning of Pretrained Transformer via Differential Inclusions (ICLR 2025 · similarity 0.87)
- ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation (NeurIPS 2025 · similarity 0.87)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.87)
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.
