New WTF algorithm fine-tunes generative models with optimal transport

Researchers have introduced Wasserstein-Tilted Flow Maps (WTF), a novel simulation-free reinforcement learning algorithm for fine-tuning flow-based generative models. Unlike existing methods that rely on KL-regularized reward maximization, WTF utilizes an optimal transport regularizer derived from the pre-trained drift. This approach directly transports individual samples toward higher rewards, framing the problem as a deterministic optimal control problem on the flow. Experiments on ImageNet-256 and text-to-image generation demonstrate that WTF achieves superior reward alignment and comparable or better diversity than baseline methods, while significantly reducing computational requirements. AI

IMPACT Introduces a more computationally efficient method for fine-tuning generative models, potentially accelerating research and development in generative AI.

RANK_REASON Academic paper detailing a new algorithm and experimental results. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• Abbas Mamedov
• flow-based generative model
• Hugging Face
• ImageNet-256
• KL-regularized reward maximization
• optimal transport
• reinforcement learning
• Wasserstein-Tilted Flow Maps

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →