Research map / Image generation / Fast diffusion sampling
SlimFlow: Training Smaller One-Step Diffusion Models with Rectified Flow
Yuanzhi Zhu, Xingchao Liu, Qiang Liu
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 annealing schedule to reflow (ablation) | small 2-rectified flow · CIFAR10 (RK45) | FID: 5.84 -> 5.46 | Improved |
| add horizontal-flip data augmentation of reflow pairs (ablation) | small 2-rectified flow · CIFAR10 (RK45) | FID: 5.84 -> 5.06 | Improved |
| combine annealing and H-Flip (ablation) | small 2-rectified flow · CIFAR10 (RK45) | FID: 5.84 -> 4.51 | Improved |
| use any annealing beta schedule instead of beta=0 (ablation) | Annealing Reflow · CIFAR10 | FID: 5.06 -> 4.51 | Improved |
| add 2-step regularization L_2-step with l2 loss (ablation) | one-step distillation · CIFAR10 (15.7M) | FID: 7.90 -> 7.30 | Improved |
| add 2-step regularization L_2-step with LPIPS loss (ablation) | one-step distillation · CIFAR10 (15.7M) | FID: 6.43 -> 5.81 | Improved |
| initialize distilled student from 2-rectified flow weights (ablation) | one-step distillation · CIFAR10 (15.7M) | FID: 9.19 -> 7.90 | Improved |
| distill from 2-rectified flow data instead of 1-rectified flow data (ablation) | one-step distillation · CIFAR10 (15.7M) | FID: 8.98 -> 7.90 | Improved |
| use LPIPS instead of l2 distillation loss (ablation) | one-step distillation · CIFAR10 (15.7M) | FID: 7.90 -> 6.43 | Improved |
| use EDM teacher instead of 1-Rectified Flow teacher (ablation) | SlimFlow 15.7M · CIFAR10 | FID: 5.81 -> 5.02 | Improved |
| use fixed larger beta (up to 0.5) during Annealing Reflow (ablation) | Annealing Reflow · 15M model | one-step sample quality | Hurt |
| halve parameters (80.7M to 44.7M) with similar MACs on ImageNet (ablation) | SlimFlow · ImageNet 64x64 | FID: 12.34 -> 12.52 | No clear effect |
Closest papers by meaning
- Shortcutting Pre-trained Flow Matching Diffusion Models is Almost Free Lunch (NeurIPS 2025 · similarity 0.90)
- Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow (ICLR 2025 · similarity 0.90)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.90)
- One Step Diffusion via Shortcut Models (ICLR 2025 · similarity 0.90)
- Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction (NeurIPS 2025 · similarity 0.90)
- Efficient Dataset Distillation via Minimax Diffusion (CVPR 2024 · 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.
