Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards

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 .

Generative AI can now create images, videos, articles, or even control robots. In real-world tasks, we often want these large models to satisfy several instructions at once — for example, "make my face in this photo look younger AND make my hair straight," or "go to the kitchen and get me an apple, but don't pass by the living room." However, when generative AI tries to compose multiple instructions, it often produces hallucinated outputs — for instance, the edited photo becomes distorted and unrealistic. We developed a method that automatically detects when such failures occur during generation and corrects them on the fly. The idea is intuitive: when all instructions "vote" for the same direction, the result satisfies every instruction while preserving fidelity. Surprisingly, this simple principle achieves strong empirical results and aligns with our theoretical analysis of why existing methods fail. Across tasks ranging from image editing to robotic control, our method satisfies multiple constraints more reliably than existing approaches while using much less compute. This makes generative AI more practical for real-world applications.