Research map / Image generation / Text-to-image diffusion
It's Never Too Late: Noise Optimization for Collapse Recovery in Trained Diffusion Models
Anne Harrington, A. Sophia Koepke, Shyamgopal Karthik, Trevor Darrell, Alexei A. Efros
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 |
|---|---|---|---|
| optimize initial noise to maximize diversity (DINO+CLIP) instead of i.i.d. noise | SDXL-Turbo · GenEval, white noise | DINO diversity: 0.588 -> 0.784 | Improved |
| noise optimization versus Parmar et al. candidate selection | SDXL-Turbo · GenEval, white noise | DINO diversity: 0.705 -> 0.784 | Improved |
| optimize noise with DPP diversity and HPSv2 reward | SDXL-Turbo · GenEval | Vendi score: 2.000 -> 4.000 | Improved |
| optimize noise with DPP diversity and HPSv2 reward | SDXL-Turbo · GenEval | HPSv2: 0.284 -> 0.292 | No clear effect |
| optimize noise with DPP and HPSv2 | Flux.1 [schnell] · GenEval, white noise | Vendi score: 2.013 -> 2.753 | Improved |
| optimize noise with DPP and HPSv2 | Flux.1 [schnell] · GenEval, white noise | HPSv2: 0.304 -> 0.293 | Hurt |
| use set-level objectives (DPP, Vendi) instead of pairwise DINO objective (ablation) | noise optimization · SDXL-Turbo human preference study | win rate | Improved |
| pink noise (alpha=0.2) initialization instead of white (ablation) | i.i.d. sampling, SDXL-Turbo · GenEval | DINO diversity: 0.588 -> 0.642 | Improved |
| pink noise (alpha=0.2) initialization instead of white (ablation) | noise optimization, SDXL-Turbo · GenEval | DINO diversity: 0.784 -> 0.786 | Improved |
| pink noise initialization for Parmar et al. (ablation) | SDXL-Turbo · GenEval | DINO diversity: 0.705 -> 0.749 | Improved |
| increase pink noise exponent alpha above 0.2 (ablation) | noise initialization · GenEval | image quality | Hurt |
| pink noise initialization | noise optimization · GenEval | iterations to beat Parmar et al.: 9 -> 12 iterations (pool 64, white vs alpha=0.2 pink) | Hurt |
Closest papers by meaning
- InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization (CVPR 2024 · similarity 0.89)
- Adjusting Initial Noise to Mitigate Memorization in Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Test-Time Scaling of Diffusion Models via Noise Trajectory Search (NeurIPS 2025 · similarity 0.88)
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.87)
- Fast constrained sampling in pre-trained diffusion models (NeurIPS 2025 · similarity 0.87)
- CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Sampling (ICLR 2024 · similarity 0.87)
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.
