RewardFlow: Generate Images by Optimizing What You Reward
Onkar Susladkar, Dong-Hwan Jang, Tushar Prakash, Adheesh Juvekar, Vedant Shah + 5 more
TLDR
RewardFlow is an inversion-free framework that steers diffusion models using multi-reward Langevin dynamics for state-of-the-art image editing and generation.
Key contributions
- Introduces RewardFlow, an inversion-free framework for steering diffusion and flow-matching models.
- Unifies complementary differentiable rewards for semantic alignment, perceptual fidelity, and human preference.
- Presents a novel VQA-based reward for fine-grained language-vision semantic supervision.
- Employs a prompt-aware adaptive policy to dynamically coordinate diverse reward objectives.
Why it matters
This paper introduces a novel method to guide image generation models more effectively by combining various reward signals. It significantly improves image editing and compositional generation quality, offering a powerful new tool for creative AI applications. The VQA-based reward is particularly innovative for fine-grained control.
Original Abstract
We introduce RewardFlow, an inversion-free framework that steers pretrained diffusion and flow-matching models at inference time through multi-reward Langevin dynamics. RewardFlow unifies complementary differentiable rewards for semantic alignment, perceptual fidelity, localized grounding, object consistency, and human preference, and further introduces a differentiable VQA-based reward that provides fine-grained semantic supervision through language-vision reasoning. To coordinate these heterogeneous objectives, we design a prompt-aware adaptive policy that extracts semantic primitives from the instruction, infers edit intent, and dynamically modulates reward weights and step sizes throughout sampling. Across several image editing and compositional generation benchmarks, RewardFlow delivers state-of-the-art edit fidelity and compositional alignment.
📬 Weekly AI Paper Digest
Get the top 10 AI/ML arXiv papers from the week — summarized, scored, and delivered to your inbox every Monday.