Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards
About
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.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Guided Flow Matching | Synthetic Dataset 10k samples 2-dimensional Mixture of Gaussians | Posterior Coverage94.6 | 20 | |
| Robotic Manipulation | ManiSkill2 StackCube (static obstacles) | Violations0.1 | 8 | |
| Robotic Manipulation | ManiSkill2 StackCube (hybrid composition) | Violation0.4 | 8 | |
| Robotic Planning | Maze2D static goal (100 samples) | Safety100 | 6 | |
| Robotic Planning | Maze2D dynamic obstacles (100 samples) | Safety Score96 | 6 | |
| Robotic Planning | Maze2D hybrid composition (100 samples) | Safety Score95 | 6 | |
| Robotic Planning | Maze2D static obstacles (100 samples) | Safety100 | 6 | |
| Text-Guided Image Manipulation | CelebA-HQ (test) | LPIPS0.226 | 5 | |
| PickCube | ManiSkill2 PickCube (static obstacles) | Violation Count0.00e+0 | 4 | |
| PickCube | ManiSkill2 PickCube (static goal) | Violation0.1 | 4 |