Research map / Image generation / Fast diffusion sampling
Glance: Accelerating Diffusion Models with 1 Sample
Zhuobai Dong, Rui Zhao, songjie wu, Suyang Hou, Junchao Yi, Zhengyuan Yang, Lijuan Wang, Alex Jinpeng 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 |
|---|---|---|---|
| add Fast-LoRA for late steps (ablation) | Slow3 + Base5 · OneIG-Bench (Qwen-Image) | Alignment: 0.805 -> 0.849 | Improved |
| add Slow-LoRA for early steps (ablation) | Base3 + Fast5 · OneIG-Bench (Qwen-Image) | Alignment: 0.747 -> 0.849 | Improved |
| use two phase-specific LoRAs instead of a single LoRA at identical timesteps (ablation) | Single (identical) · OneIG-Bench (Qwen-Image) | Alignment: 0.702 -> 0.849 | Improved |
| use phase-aware Slow-Fast instead of single uniform LoRA (ablation) | Single (uniform) · OneIG-Bench (Qwen-Image) | Alignment: 0.621 -> 0.849 | Improved |
| scale training set from 1 to 10 samples (ablation) | Glance (Qwen 10x2) · OneIG-Bench (Qwen-Image) | Alignment: 0.868 -> 0.874 | No clear effect |
| scale training set from 1 to 100 samples (ablation) | Glance (Qwen 10x2) · OneIG-Bench (Qwen-Image) | Alignment: 0.868 -> 0.876 | No clear effect |
| scale training set from 1 to 10 samples (ablation) | Glance (Qwen 10x2) · OneIG-Bench (Qwen-Image) | Style: 0.421 -> 0.414 | Hurt |
| increase LoRA timesteps from Slow3+Fast5 to Slow5+Fast5 (ablation) | Glance (Qwen) · OneIG-Bench (Qwen-Image) | Text: 0.692 -> 0.734 | Improved |
| increase LoRA timesteps from Slow5+Fast5 to Slow5+Fast10 (ablation) | Glance (Qwen) · OneIG-Bench (Qwen-Image) | Text: 0.734 -> 0.813 | Improved |
| train on text-rich Bookstore sample instead of Fox (ablation) | Glance (Qwen 10x2) · OneIG-Bench (Qwen-Image) | Alignment: 0.868 -> 0.797 | Hurt |
| train on out-of-domain real-world image instead of Fox (ablation) | Glance (Qwen 10x2) · OneIG-Bench (Qwen-Image) | Alignment: 0.868 -> 0.857 | No clear effect |
| train on pure Gaussian noise (ablation) | Glance · single-sample training | image quality | Hurt |
Closest papers by meaning
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.92)
- Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.91)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.91)
- Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network (NeurIPS 2026 · similarity 0.91)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.91)
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.90)
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.
