This post introduces the DRaFT+ algorithm for fine-tuning generative text-to-image diffusion models. The algorithm enhances the alignment between text and generated images, prevents mode collapse, and improves generation diversity. The DRaFT+ algorithm is accessible through the NeMo-Aligner library on GitHub.

7m read time From developer.nvidia.com
Post cover image
Table of contents
Direct reward fine-tuning (DRaFT)DRaFT+Results of DRaFT+ trainingSummary

Sort: