Research map / Image generation / Fast diffusion sampling
FastFLUX: Pruning FLUX with Block-wise Replacement and Sandwich Training
Fuhan Cai, Yong Guo, Jie Li, Wenbo Li, Jian Chen, Xiangzhong Fang
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 Sandwich Training (LoRA on neighbouring blocks) (ablation) | FastFLUX-10% · HPS v2 | HPS v2 score: 27.99 -> 28.58 | Improved |
| add BRLL linear replacement (ablation) | FastFLUX · HPS v2 | HPS v2 score: 28.25 -> 28.58 | Improved |
| use 2 LoRA-enhanced layers instead of 0 (ablation) | Sandwich Training · HPS v2 | HPS v2 score: +0.60 | Improved |
| use 4 LoRA-enhanced layers instead of 2 (ablation) | Sandwich Training · HPS v2 | HPS v2 score: +0.03 | No clear effect |
| use importance-guided order instead of Start2End (ablation) | FastFLUX · HPS v2 | HPS v2 score: 24.63 -> 28.58 | Improved |
| use importance-guided order instead of End2Start (ablation) | FastFLUX · HPS v2 | HPS v2 score: 28.15 -> 28.58 | Improved |
| replace L1-norm pruning with FastFLUX at similar latency (12% vs 10%) | FLUX.1-dev · HPS v2 | HPS v2 score: 28.20 -> 28.58 | Improved |
| replace L1-norm pruning with FastFLUX at similar latency (27% vs 20%) | FLUX.1-dev · HPS v2 | HPS v2 score: 27.17 -> 27.76 | Improved |
| prune 10% of blocks with FastFLUX | FLUX.1-dev · HPS v2 | HPS v2 score: 28.32 -> 28.58 | Improved |
| replace L1-norm pruning with FastFLUX at similar latency (27% vs 20%) | FLUX.1-dev · GenEval | overall score: 0.47 -> 0.53 | Improved |
| prune 10% with FastFLUX vs L1 pruning 12% | FLUX.1-dev · GenEval | overall score: 0.60 -> 0.60 | No clear effect |
| increase pruning ratio from 10% to 30% | FastFLUX · HPS v2 | HPS v2 score: 28.58 -> 26.70 | Hurt |
Closest papers by meaning
- TinyFusion: Diffusion Transformers Learned Shallow (CVPR 2025 · similarity 0.88)
- Not All Prompts Are Made Equal: Prompt-based Pruning of Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.88)
- ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation (NeurIPS 2025 · similarity 0.87)
- F³-Pruning: A Training-Free and Generalized Pruning Strategy towards Faster and Finer Text-to-Video Synthesis (AAAI 2024 · similarity 0.87)
- SANA 1.5: Efficient Scaling of Training-Time and Inference-Time Compute in Linear Diffusion Transformer (ICML 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.
