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

2026-06-25 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, medium

Summary

A new method called Feature Self-Guidance addresses diversity collapse in pretrained flow models, which often generate similar samples under identical conditions. This efficient, training-free self-guidance mechanism, detailed in paper 2606.27371 by R. Venkatesh Babu et al., integrates as a plug-and-play module with only marginal inference cost. It operates by dispersing the model's internal features during batch generation, followed by a manifold regularization step that projects these features back onto the data manifold. This ensures diverse generation without sacrificing alignment to input conditions. Experiments confirm significant improvements in diversity and fidelity across various conditional flow models, including multi-step and few-step text-to-image, depth-to-image, and reference image generation tasks.

Key takeaway

For Machine Learning Engineers developing or deploying conditional flow models, if you are struggling with diversity collapse, consider integrating Feature Self-Guidance. This training-free, plug-and-play module offers significant diversity improvements with only marginal inference cost, avoiding the overhead of external reward models. You can apply it across various tasks like text-to-image or depth-to-image generation to ensure diverse outputs while maintaining fidelity.

Key insights

Feature Self-Guidance efficiently mitigates diversity collapse in flow models by dispersing internal features and regularizing them to the data manifold.

Principles

• Diversity collapse is a key challenge in flow models.
• External reward models add significant inference overhead.
• Feature dispersion can enhance generative diversity.

Method

Disperse internal features during batch generation, then apply manifold regularization to project dispersed features back onto the data manifold, ensuring diversity and alignment.

In practice

• Integrate as a plug-and-play module.
• Apply to text-to-image generation.
• Use for depth-to-image tasks.

Topics

• Flow Models
• Diversity Collapse
• Feature Self-Guidance
• Manifold Regularization
• Text-to-Image Generation
• Conditional Generation

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Initialization is Half the Battle: Generating Diverse Images from a Guidance Potential Posterior — Initializing generative models from a guidance potential posterior, rather than standard Gaussian noise, significantly enhances output diversity and quality.
• Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards — Gradient misalignment causes off-manifold drift when composing multiple constraints in guided sampling for flow models.
• Consistency-Preserving Diverse Video Generation — A new flow-matching framework improves video diversity and temporal consistency using latent-space gradient regulation.
• From Collapse to Improvement: Statistical Perspectives on the Evolutionary Dynamics of Iterative Training on Contaminated Sources — Model collapse in iterative training can be avoided with fresh data and appropriate sample sizes.
• CFG-Ctrl: Control-Based Classifier-Free Diffusion Guidance — CFG can be reinterpreted as a control system, improving stability and semantic alignment with nonlinear methods.
• Flow-Direct: Feedback-Efficient and Reusable Guidance for Flow Models via Non-Parametric Guidance Field — Flow-Direct uses a persistent, non-parametric guidance field to efficiently steer generative models with black-box rewards.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.