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 generate impressive images from text or image prompts, but they suffer from diversity collapse: multiple samples generated under the same conditioning tend to converge to similar outputs. Existing remedies are limited—latent guidance offers only modest gains, while sample selection depends on external reward models that add significant inference-time overhead. This paper introduces an efficient, training-free self-guidance mechanism that addresses diversity collapse directly. The method spreads the flow model's internal features apart during batch generation via feature self-guidance, combined with manifold regularization to ensure diverse generations without sacrificing alignment with the input condition. The approach requires no retraining or external models, making it a lightweight solution for improving sample diversity in pretrained flow models. Paper: arXiv 2606.27371, authored by Pradhaan S Bhat, Rishubh Parihar, and Abhijnya Bhat (CV, June 2026).

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. The approach disperses the flow model's internal features during batch generation through feature self-guidance, and employs manifold regularization to ensure diverse generation without sacrificing alignment with the input conditioning.

Key Points

Problem: pretrained flow models produce near-identical samples when queried repeatedly with the same prompt or condition.
Prior fixes fall short: latent guidance is weakly effective; sample selection adds costly external reward-model inference.
Proposed solution: a training-free, self-guidance mechanism operating on the model's internal features during batch generation.
Feature self-guidance spreads internal features apart across the batch to encourage output diversity.
Manifold regularization keeps generated samples diverse while preserving conditioning alignment.
Requires no retraining and no external reward models, keeping inference overhead low.

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