Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards
The paper presents a novel method called Conflict-Aware Additive Guidance ($g^ ext{car}$) aimed at improving flow models under compositional rewards. This method addresses issues related to off-manifold drift caused by gradient misalignment when multiple constraints are applied simultaneously. The authors validate their approach across various domains, demonstrating enhanced generation fidelity with minimal computational resources.
- ▪The proposed method actively detects and resolves gradient conflicts to rectify off-manifold drift.
- ▪Existing methods often struggle with composing multiple constraints, leading to inaccuracies in generated outputs.
- ▪The validation of $g^ ext{car}$ spans synthetic datasets, image editing, and generative decision-making.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.20758 (cs) [Submitted on 20 May 2026] Title:Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards Authors:Xuehui Yu, Fucheng Cai, Meiyi Wang, Xiaopeng Fan, Harold Soh View a PDF of the paper titled Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards, by Xuehui Yu and 4 other authors View PDF HTML (experimental) 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.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
