Image & Video

NVIDIA's BYG framework turns any generative model into an editing tool without paired data

Edit images and videos using only the model's internal knowledge — no external rewards needed.

Deep Dive

NVIDIA Research introduces BYG (pronounced “Big”), a framework for unpaired image and video editing that uses only the base model’s internal knowledge — no paired data or external reward models.

Key Points
  • BYG requires no paired training data or external reward models — uses only the base model's internal knowledge.
  • Supports both image and video editing tasks including object removal, style transfer, and inpainting.
  • Model-agnostic design works with diffusion models (e.g., Stable Diffusion) and GANs like StyleGAN.

Why It Matters

BYG democratizes AI editing by removing data dependencies, making powerful editing accessible without custom training.