Feature Self-Guidance: Mitigating Diversity Collapse in Pretrained Flow Models (arXiv 2606.27371)

Forum topic · 小凯 · 2026-06-27

Summary

This paper introduces a training-free, efficient self-guidance mechanism to address diversity collapse in state-of-the-art flow models, which tend to produce nearly identical images when generating multiple samples under the same text or image conditioning. The authors, Pradhaan S Bhat, Rishubh Parihar, and Abhijnya Bhat, propose Feature Self-Guidance, which disperses the model's internal features across samples during batch generation, combined with manifold regularization to ensure diverse outputs without sacrificing alignment with the input condition. Compared to existing approaches—latent guidance, which has limited effectiveness, and sample selection using external reward models, which incurs significant inference-time overhead—the proposed method requires no additional training and adds minimal inference cost. The work was published on arXiv (2606.27371) and is categorized under computer vision. It is relevant to researchers working on generative models, flow-based diffusion, and diversity-preserving sampling techniques.

Paper Overview

Field: Computer Vision (CV)
Authors: Pradhaan S Bhat, Rishubh Parihar, Abhijnya Bhat
Published: 2026-06-27
arXiv: 2606.27371

Abstract (from the paper)

Key Contributions

Feature Self-Guidance: A training-free self-guidance mechanism that disperses the flow model's internal features across samples during batch generation, encouraging diverse outputs.
Manifold Regularization: Ensures diversity while preserving alignment with the input conditioning, avoiding drift away from the prompt.
Efficiency: Unlike reward-model-based sample selection, the method incurs minimal inference-time overhead and requires no additional training.

Context

Existing approaches to diversity collapse either use latent guidance (with limited effectiveness) or sample selection via external reward models (with significant inference-time cost). This work positions Feature Self-Guidance as an efficient middle ground for pretrained flow models.

--- *Auto-collected on 2026-06-27*