Research map / Image generation / Fast diffusion sampling
Accelerating Image Generation with Sub-path Linear Approximation Model
Chen Xu, Tianhui Song, Weixin Feng, Xubin Li, Tiezheng Ge, Bo Zheng, Limin 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 sub-path linear approximation (SPLAM) instead of consistency distillation (LCM) | SDv1.5 distillation, 4 steps · MSCOCO2014-30k (w=8) | FID: 14.53 -> 13.39 | Improved |
| use SPLAM instead of LCM at 2 steps | SDv1.5 distillation · MSCOCO2014-30k (w=3) | FID: 14.29 -> 12.31 | Improved |
| use SPLAM instead of LCM at 4 steps | SDv1.5 distillation · MSCOCO2017-5k (w=3) | FID: 21.41 -> 20.77 | Improved |
| use SPLAM instead of LCM with SDv2.1-base on LAION test set | SDv2.1-base distillation · LAION-Aesthetics-6+ (1 step) | FID: 32.41 -> 32.64 | No clear effect |
| use SPLAM instead of LCM with SDv2.1-base on COCO-30k zero-shot | SDv2.1-base distillation · COCO-30k (1 step) | FID: 43.87 -> 40.52 | Improved |
| add multiple estimation strategy to allow large skipping step size k (ablation) | SPLAM · SDv1.5, w=8, skipping step size ablation | FID | Improved |
| increase skipping step size k without sub-path linear ODE (LCM) (ablation) | LCM · SDv1.5, w=8, skipping step size ablation | FID | Hurt |
| raise skipping step size from 20 to 100 with multiple estimation (ablation) | SPLAM · SDv1.5 training cost | training cost: about 6 A100 GPU days vs about 16 for LCM | Improved |
| use sub-path linear ODE objective instead of LCM objective (ablation) | SPLAM · COCO samples, consistency mapping error delta | mapping error: about 10% reduction at t=800 | Improved |
| increase guidance scale w (ablation) | SPLAM · SDv1.5 | FID / CLIP Score | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- UFOGen — UFOGen: You Forward Once Large Scale Text-to-Image Generation via Diffusion GANs (CVPR 2024) · better on 1 of 1 shared result
- InstaFlow-1.7B — InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024) · 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
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.89)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.88)
- Simple Distillation for One-Step Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.88)
- Reducing Spatial Fitting Error in Distillation of Denoising Diffusion Models (AAAI 2024 · similarity 0.88)
- One-step Diffusion with Distribution Matching Distillation (CVPR 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.
