Windowed Minority Guidance: Preliminary Evidence for Timestep-Localized Effects in Diffusion Denoising

Vittoria Lanzo · Independent Researcher

Extended abstract, EEML 2026

---

Minority guidance (Um et al., 2024) steers diffusion sampling toward under-represented data regions via a classifier gradient at every denoising timestep. We test whether this effect is timestep-localized, comparing three equal-thirds windows (early: t∈[0,333), mid: t∈[333,667), late: t∈[667,1000)) against full-chain and no-guidance baselines over 50 seeds on LSUN Bedroom (250 runs). Mid recovers 0.456 of full-chain loss reduction; Wilcoxon tests confirm effects for mid, early, full vs. baseline (all p < 0.001), with late vs. baseline narrowly missing correction (p = 0.0072 vs. α/7 = 0.00714). All windows produce detectable effects: preliminary evidence for mid-chain dominance without strict localization. All measurements use guidance_scale=1.0; Um et al. (2024) report at 3.5, and generalization to that scale requires validation.

---

Timestep Convention. Window names follow the numerical magnitude of the timestep index t, not the chronological order of the DDPM reverse process. "Early" (t∈[0,333)) corresponds to low-noise timesteps (fine-detail refinement); "late" (t∈[667,1000)) to high-noise timesteps (coarse-structure formation). Concretely:

Window	t range	Noise level	Denoising phase
early	[0, 333)	low	fine-detail refinement
mid	[333, 667)	intermediate	semantic layout
late	[667, 1000)	high	coarse structure

---

Results Summary

Condition	Mean Loss	Win Rate	Relative Effect
baseline	11.978	N/A	0.000
minority-early	11.129	50/50	0.162
minority-mid	9.591	39/50	0.456
minority-late	10.717	34/50	0.241
minority-full	6.741	44/50	1.000

Mid, early, and full guidance versus baseline all reach p < 0.001 (Wilcoxon signed-rank on 50 paired seeds, Bonferroni threshold α/7 ≈ 0.00714); late versus baseline (p = 0.0072) narrowly misses correction.

---

Extended abstract

📄 WMG.preliminary.pdf · WMG.preliminary.tex (LaTeX source)

Reproduce

🔗 windowed-minority-guidance-experiment

Run at guidance_scale=1.0. Um et al. (2024) report results at guidance_scale=3.5; the windowed effects measured here may not generalize to the published scale and require validation there.

---

Keywords

minority guidance, classifier-guided diffusion, windowed guidance, DDPM, LSUN Bedroom, inference-time guidance, timestep localization, guided diffusion sampling

---

Citation

@misc{lanzo2026windowed,
  title  = {Windowed Minority Guidance: Preliminary Evidence for Timestep-Localized Effects in Diffusion Denoising},
  author = {Vittoria Lanzo},
  year   = {2026},
  note   = {Extended abstract, EEML 2026},
  url    = {https://github.com/VittoriaLanzo/windowed-minority-guidance}
}

References

@inproceedings{dhariwal2021diffusion,
  title     = {Diffusion Models Beat {GAN}s on Image Synthesis},
  author    = {Prafulla Dhariwal and Alexander Nichol},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2021}
}

@inproceedings{ho2020denoising,
  title     = {Denoising Diffusion Probabilistic Models},
  author    = {Jonathan Ho and Ajay Jain and Pieter Abbeel},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2020}
}

@inproceedings{meng2022sdedit,
  title     = {{SDEdit}: Guided Image Synthesis and Editing with Stochastic Differential Equations},
  author    = {Chenlin Meng and Yutong He and Yang Song and Jiaming Song and Jiajun Wu and Jun-Yan Zhu and Stefano Ermon},
  booktitle = {International Conference on Learning Representations},
  year      = {2022}
}

@inproceedings{um2024dont,
  title     = {Don't Play Favorites: Minority Guidance for Diffusion Models},
  author    = {Soobin Um and Suhyeon Lee and Jong Chul Ye},
  booktitle = {International Conference on Learning Representations},
  year      = {2024}
}