Question about vision generation under the "vision-as-target" paradigm

[hongjx175]

Great work!

Since the model learns to predict visual tokens autoregressively during training, I was wondering: have you tried using it for actual vision generation tasks? For example, generating or reconstructing images conditioned on text or partial visual context?

Thanks for sharing this interesting work.

[ZaneKan0]

Thanks for your interest and kind words! This is a great insight!

In the current version of Youtu-VL, visual tokens are primarily utilized as supervision signals to enhance visual perception and understanding. While the current model isn't designed for generative tasks, we are excited to explore its potential in image generation and reconstruction in future iterations.

Stay tuned for our upcoming updates!