Don't Settle at the Mode: Training-Free Feature Self-Guidance Mitigates Diversity Collapse in Flow Models

Forum topic · 小凯 · 2026-06-27

Summary

A new arXiv paper (2606.27371) by Pradhaan S Bhat, Rishubh Parihar, and Abhijnya Bhat addresses diversity collapse in state-of-the-art flow generative models. When flow models generate multiple samples under identical text or image conditioning, outputs tend to cluster around a single mode, reducing variety. Existing solutions fall into two camps: latent guidance, which offers limited effectiveness, and sample selection using external reward models, which adds significant inference-time overhead. The authors propose an efficient, training-free self-guidance mechanism that operates directly on the model's internal features. During batch generation, the method disperses the flow model's internal features across samples, while a manifold regularization term ensures generated images remain diverse without sacrificing alignment with the input conditioning. The approach requires no retraining and avoids external reward models, making it a lightweight inference-time fix for diversity collapse in flow-based text-to-image and image-to-image generation.

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 — outputs converge toward a single mode instead of spanning the full distribution.

Limitations of prior approaches

Latent guidance: limited effectiveness in restoring diversity.
Sample selection: relies on external reward models, incurring significant inference-time overhead.

Proposed method

training-free self-guidance mechanism

Feature self-guidance: during batch generation, the flow model's internal features are dispersed across samples in the batch, encouraging distinct outputs.
Manifold regularization: ensures diverse generation without sacrificing alignment with the input conditioning.

The method requires no retraining and no external reward models, making it a lightweight inference-time solution for improving sample diversity in flow-based generative models.

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