Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AdaptDiff: Adaptive Guidance in Diffusion Models for Diverse and Identity-Consistent Face Synthesis (Student Abstract)

Eduarda Caldeira, Tahar Chettaoui, Naser Damer, Fadi Boutros

Abstract

Diffusion models conditioned on identity embeddings enable the generation of synthetic face images that consistently preserve identity across multiple samples. Recent work has shown that introducing an additional negative condition through classifier-free guidance during sampling provides a mechanism to suppress undesired attributes, thus improving inter-class separability. Building on this insight, we propose a dynamic weighting scheme for the negative condition that adapts throughout the sampling trajectory. This strategy leverages the complementary strengths of positive and negative conditions at different stages of generation, leading to more diverse yet identity-consistent synthetic data.

AdaptDiff: Adaptive Guidance in Diffusion Models for Diverse and Identity-Consistent Face Synthesis (Student Abstract)

Abstract

Diffusion models conditioned on identity embeddings enable the generation of synthetic face images that consistently preserve identity across multiple samples. Recent work has shown that introducing an additional negative condition through classifier-free guidance during sampling provides a mechanism to suppress undesired attributes, thus improving inter-class separability. Building on this insight, we propose a dynamic weighting scheme for the negative condition that adapts throughout the sampling trajectory. This strategy leverages the complementary strengths of positive and negative conditions at different stages of generation, leading to more diverse yet identity-consistent synthetic data.

Paper Structure

This paper contains 18 sections, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Results
  4. $w$ & $w_{max}$ Ablation Study:
  5. Identity Separability:
  6. Face Recognition:
  7. Conclusion
  8. Ethics Statement
  9. Acknowledgments
  10. Supplementary Material
  11. Experimental Setup
  12. Pre-Trained Diffusion Models:
  13. Sampling Methods:
  14. Identity Separability Evaluation:
  15. AdaptDiff FR Training Setup:
  16. ...and 3 more sections

Figures (3)

  • Figure 1: Histograms of genuine and impostor score distributions of SOTA baselines and AdaptDiff trained on FFHQ.
  • Figure 2: $w$ as a function of the sampling time step, $t$. Blue: IDiff-Face fixes $w=0$. Red: NegFaceDiff fixes $w=0.5$. Yellow: AdaptDiff adapts $w$ during sampling ($w=1-t/T$).
  • Figure 3: Histograms of genuine and impostor score distributions of SOTA baselines and AdaptDiff trained on C-WF.