Reward-guided Fine-Tuning of One-Step Generative Models via Wasserstein Gradient Flow

Name: Reward-guided Fine-Tuning of One-Step Generative Models via Wasserstein Gradient Flow
Uploaded: 2026-09-08T07:03:13.279886+00:00
Duration: 280 s
Description: The episode discusses a method using Wasserstein Gradient Flow to train generative models by guiding their output toward a desired target defined by a reward function. The innovation lies in using Unbalanced Optimal Transport to handle complex, non-differentiable rewards. This approach improves reward alignment, stability, and performance on tasks like ImageNet.

The episode discusses a method using Wasserstein Gradient Flow to train generative models by guiding their output toward a desired target defined by a reward function. The innovation lies in using Unbalanced Optimal Transport to handle complex, non-differentiable rewards. This approach improves reward alignment, stability, and performance on tasks like ImageNet.

Open the full episode