Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards

2026-05-20

AI Summary

This work addresses the degradation in generation quality caused by multi-constraint compositional guidance, which often drives generative models away from the true data manifold. The authors propose g^car, a lightweight and learnable conflict-aware additive guidance method that dynamically detects and mitigates gradient conflicts among multiple objectives during inference, effectively correcting manifold drift in flow models under composite rewards. The study is the first to systematically uncover the connection between gradient misalignment and manifold drift in multi-constraint guidance and introduces a dynamic conflict-aware mechanism to achieve efficient alignment. Experiments demonstrate that the proposed approach significantly improves generation fidelity across synthetic data, image editing, and decision-making tasks, outperforming existing baselines while maintaining low computational overhead.

Abstract

Inference-time guided sampling steers state-of-the-art diffusion and flow models without fine-tuning by interpreting the generation process as a controllable trajectory. This provides a simple and flexible way to inject external constraints (e.g., cost functions or pre-trained verifiers) for controlled generation. However, existing methods often fail when composing multiple constraints simultaneously, which leads to deviations from the true data manifold. In this work, we identify root causes of this off-manifold drift and find that the approximation error scales severely with gradient misalignment. Building on these findings, we propose Conflict-Aware Additive Guidance ($g^\text{car}$), a lightweight and learnable method, which actively rectifies off-manifold drift by dynamically detecting and resolving gradient conflicts. We validate $g^\text{car}$ across diverse domains, ranging from synthetic datasets and image editing to generative decision-making for planning and control. Our results demonstrate that $g^\text{car}$ effectively rectifies off-manifold drift, surpassing baselines in generation fidelity while using light compute. Code is available at https://github.com/yuxuehui/CAR-guidance.

Problem

• compositional rewards
• off-manifold drift
• gradient conflict
• guided sampling
• flow models

Innovation

• Conflict-Aware Guidance
• Flow Models
• Off-Manifold Drift
• Gradient Alignment
• Compositional Rewards