Research map / Image generation / Text-to-image diffusion
Golden Noise for Diffusion Models: A Learning Framework
Zikai Zhou, Shitong Shao, Lichen Bai, Shufei Zhang, Zhiqiang Xu, Bo Han, Zeke Xie
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 singular value prediction (ablation) | NPNet (SDXL) · Pick-a-Pic | ImageReward: 49.03 -> 65.01 | Improved |
| add residual prediction (ablation) | NPNet (SDXL) · Pick-a-Pic | ImageReward: 63.05 -> 65.01 | Improved |
| add data selection with human preference model (ablation) | NPNet (SDXL) · Pick-a-Pic | ImageReward: 62.91 -> 65.01 | Improved |
| replace standard noise with NPNet golden noise | SDXL · Pick-a-Pic | ImageReward: 58.01 -> 65.01 | Improved |
| use NPNet instead of plain re-denoise sampling (Inversion) | SDXL · DrawBench | HPSv2: 26.91 -> 27.14 | Improved |
| apply NPNet to DreamShaper-xl-v2-turbo | DreamShaper-xl-v2-turbo · HPD v2 | HPSv2: 30.96 -> 34.08 | Improved |
| apply plain re-denoise sampling (Inversion) without learning | DreamShaper-xl-v2-turbo · DrawBench | HPSv2: 30.39 -> 30.10 | Hurt |
| fine-tune NPNet on 600 Hunyuan-DiT noise pairs | Hunyuan-DiT · Pick-a-Pic | ImageReward: 91.33 -> 100.82 | Improved |
| add NPNet on top of DPO | DPO on SDXL · Pick-a-Pic (10 steps) | ImageReward: 68.82 -> 75.06 | Improved |
| add NPNet on top of AYS | AYS on SDXL · Pick-a-Pic (10 steps) | ImageReward: 50.74 -> 59.50 | Improved |
| use NPNet instead of InitNO noise optimization | StableDiffusion-v1-4 · Pick-a-Pic | ImageReward: -205.66 -> -95.89 | Improved |
| raise HPSv2 filtering threshold m from 0 to 0.01 (ablation) | NPNet data selection (SDXL) · Pick-a-Pic | PickScore: 21.86 -> 21.72 | Hurt |
Closest papers by meaning
- InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization (CVPR 2024 · similarity 0.90)
- Diffusion Models With Learned Adaptive Noise (NeurIPS 2024 · similarity 0.90)
- Ambient Diffusion Omni: Training Good Models with Bad Data (NeurIPS 2025 · similarity 0.90)
- Information Theoretic Learning for Diffusion Models with Warm Start (NeurIPS 2025 · similarity 0.89)
- Fast constrained sampling in pre-trained diffusion models (NeurIPS 2025 · similarity 0.89)
- Consistent Diffusion Meets Tweedie: Training Exact Ambient Diffusion Models with Noisy Data (ICML 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.
