Paper: 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

State-of-the-art flow models can generate stunning images from text or image prompts, but they suffer from diversity collapse when multiple samples are generated under the same conditioning. Existing remedies have drawbacks: latent guidance offers limited effectiveness, while sample selection depends on external reward models that add significant inference-time overhead. This paper (arXiv 2606.27371) introduces an efficient, training-free self-guidance mechanism that mitigates diversity collapse by dispersing the flow model's internal features during batch generation. A manifold regularization term ensures diverse generation without sacrificing alignment with the input condition. The approach requires no retraining and no external reward model, making it a lightweight alternative to prior methods for improving sample diversity in pretrained flow models. Authors: Pradhaan S Bhat, Rishubh Parihar, Abhijnya Bhat. Full abstract and paper are available on arXiv.

Paper Overview

Field: 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, we introduce an efficient, training-free self-guidance mechanism to mitigate diversity collapse.

Key Points

Problem: Flow models collapse to the mode (low diversity) when producing multiple samples with identical conditioning.
Limitations of prior work: Latent guidance is only weakly effective; sample selection requires external reward models, adding heavy inference-time cost.
Proposed method: A training-free feature self-guidance mechanism that disperses the model's internal features across a batch during generation.
Manifold regularization: Ensures generated samples remain diverse while preserving alignment with the input conditioning.