Don't Settle at the Mode! Mitigating Diversity Collapse in Pretrained Flow Models via Feature Self-Guidance

Forum topic · 小凯 · 2026-06-27

Summary

This paper introduces a training-free, feature-based self-guidance mechanism that mitigates diversity collapse in pretrained flow models. State-of-the-art flow models can generate stunning images from text or image prompts, but when multiple samples are generated under the same conditioning, outputs tend to converge to a single mode, losing diversity. Existing remedies fall into two camps: latent guidance, whose effectiveness is limited, and sample selection, which relies on external reward models that add significant inference-time overhead. The proposed self-guidance approach instead leverages the model's internal features to steer sampling away from settled modes, avoiding extra reward-model costs and external dependencies. The work is in the computer vision field, authored by Pradhaan S Bhat, Rishubh Parihar, and Abhijnya Bhat, and is available on arXiv (2606.27371). The forum post includes the paper's abstract and links for further reading.

Paper Overview

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

Abstract

State-of-the-art flow models generate stunning images from text or image prompts. However, they suffer from diversity collapse when generating multiple samples under the same conditioning. Existing methods address this issue via either latent guidance, which has limited effectiveness, or sample selection, which relies on external reward models that incur significant inference-time overhead. In this work, the authors introduce an efficient, training-free self-guidance mechanism to mitigate diversity collapse.

Key Points

Problem: Pretrained flow models exhibit diversity collapse — repeated sampling under identical conditioning tends to settle at a single mode.
Limitations of prior work: Latent guidance offers limited effectiveness; sample selection depends on external reward models, adding significant inference-time overhead.
Proposed solution: An efficient, training-free self-guidance mechanism that mitigates diversity collapse without external reward models.