[paulluuk]

I work on AI solutions for a major video streaming company, and the problem with VEO3 is that it doesn't have any consistency between prompts. E.g. I can not upload a reference image of what a character looks like, and if I say in one video "the old priest bends down" and in the next video "the old priests picks up the coin", the priest will look very different between shots.

Veo3 does support image to video, so what you can do is create an image that is the start of a scene, and then use that to generate the actual scene. Unfortunately, Veo3 is really bad at this. I expect this will improve over time.

Although I'm not super excited about this Seedance model personally, I do really like that it focuses on consistency between shots. I hope this puts pressure on increased performance from Veo3 in that regard.

[Uehreka]

> I expect this will improve over time.

I’m starting to wonder if it will. There seems to be this pattern that an awesome T2V model will come out and everyone starts clamoring for an I2V model and then when the I2V version drops a couple months later it’s not as good. I’m starting to get the feeling that I2V is just intrinsically challenging in a way that makes it hard to do well at all.