Kun Ding

Leveraging the pre-trained diffusion model as a powerful and cost-effective step-level reward model to optimize the diffusion model itself directly in the noisy latent space.

Give VLM the ability to call tools to retrieve and filter information