Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards
A new method called Conflict-Aware Additive Guidance (g^car) addresses off-manifold drift in diffusion and flow models during inference-time guided sampling with multiple constraints. The approach dynamically detects and resolves gradient conflicts to maintain generation quality. Validated on synthetic datasets and image editing tasks, g^car offers a lightweight, learnable solution for controlled generation without fine-tuning.
Key facts
- arXiv:2605.20758v1 announces Conflict-Aware Additive Guidance (g^car)
- g^car rectifies off-manifold drift caused by gradient misalignment
- Method is lightweight and learnable
- Validated on synthetic datasets and image editing
- Does not require fine-tuning of base models
Entities
—