Research map / Image generation / Fast diffusion sampling
A Noise is Worth Diffusion Guidance
Donghoon Ahn, Jiwon Kang, Sanghyun Lee, Jaewon Min, Minjae Kim, Wooseok Jang, Hyoungwon Cho, Sayak Paul, SeonHwa Kim, Eunju Cha, Kyong Hwan Jin, Seungryong Kim
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 |
|---|---|---|---|
| Refine initial noise with NoiseRefine instead of Gaussian noise (no guidance in both) | Stable Diffusion 2.1 · MS-COCO 2014 val (30K prompts) | FID: 42.71 -> 11.39 | Improved |
| Refined noise without guidance versus Gaussian noise with CFG+PAG | Stable Diffusion 2.1 · MS-COCO 2014 val (30K prompts) | FID: 13.38 -> 11.39 | Improved |
| Refined noise versus guidance on Inception Score | Stable Diffusion 2.1 · MS-COCO 2014 val (30K prompts) | Inception Score: 37.64 -> 35.73 | Hurt |
| Replace guidance by refined noise to halve inference cost | Stable Diffusion 2.1 · MS-COCO 2014 val (30K prompts) | inference time (s): 2.589 -> 1.504 | Improved |
| Refine noise instead of Gaussian noise on DrawBench | Stable Diffusion 2.1 · DrawBench | PickScore: 19.67 -> 20.90 | Improved |
| Increase training denoising steps from 5 to 10 (ablation) | NoiseRefine · MS-COCO FID (ablation) | FID: 13.74 -> 13.36 | Improved |
| Add aesthetic-score filtering of training data (ablation) | NoiseRefine · MS-COCO FID (ablation) | FID: 15.27 -> 13.55 | Improved |
| Initialize noise refiner from pretrained diffusion UNet with LoRA instead of from scratch (ablation) | NoiseRefine · Filtered MS-COCO (25K steps) | FID: 37.87 -> 13.74 | Improved |
| Add multistep score distillation (detach gradients through denoiser) (ablation) | NoiseRefine training · Optimization convergence | convergence speed and image quality | Improved |
| Train noise refiner instead of LoRA-fine-tuning the denoising network (ablation) | Guidance-free T2I pipeline · Filtered MS-COCO training | convergence and quality | Improved |
| Refined noise versus Gaussian noise plus CFG in user study | Stable Diffusion 2.1 · User study (45 participants) | image quality preference (%): 46.04% -> 53.96% | No clear effect |
Closest papers by meaning
- Rectified Diffusion Guidance for Conditional Generation (CVPR 2025 · similarity 0.92)
- Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models (ICLR 2025 · similarity 0.91)
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.91)
- NoiseCtrl: A Sampling-Algorithm-Agnostic Conditional Generation Method for Diffusion Models (CVPR 2025 · similarity 0.91)
- Guiding a Diffusion Model with a Bad Version of Itself (NeurIPS 2024 · 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.
