Setup Discussion

#1
by CharlesCXK - opened

Hi, this setup is not very reasonable. Janus has not encountered image-to-image generation data during training, so it does not yet possess this capability.

@CharlesCXK this is more about assessing zero shot generalization on that task without training.

If you try with a few images, you will see that the output images almost always reflect some semantic feature of the input image. I think it's fair to say that Janus do possess this ability in some capacity, it's just not very developed.

Also visualizing the input like this could help understanding at a glance what the model "sees".

@thomasgauthier
Got it, it's an interesting experiment!

Sign up or log in to comment