New method fine-tunes generative models using Wasserstein Gradient Flow

Researchers have developed a new method for fine-tuning one-step generative models, which map noise directly to data in a single pass. This approach, termed Reward-guided Fine-Tuning via Wasserstein Gradient Flow (WGF), views one-step generators through an optimal transport lens to achieve smooth and controlled distributional evolution. The proposed training method does not require reward gradients, making it applicable to both differentiable and non-differentiable rewards, while also mitigating issues like reward hacking and mode collapse. Experiments on synthetic data, CIFAR-10, and ImageNet demonstrated improved reward alignment compared to existing methods. AI

IMPACT This new fine-tuning method could lead to more efficient and controllable generative models, potentially impacting fields that rely on image synthesis and data generation.

RANK_REASON Academic paper detailing a novel method for generative models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• CIFAR-10
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv
• ImageNet
• Influence Flower
• ScienceCast

• paper
• model release

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