Generative image synthesis animation exploration of the song titles of the WMF cd Hot Malasada. The same story line text prompt key-framing for all of these. The variation in the imagery across the different animations (as well as the flow while in specific keyframe sections) comes from how you modulate the underlying generative system.
Here's a way to think about it. The normal random U-Net latent initialization for the latent diffusion has very specific statistical properties that are very restricted compared to what they potentially could be. So if you break out of that box and try to utilize all of the potential range of that starting point U-Net latent representation you have a much bigger (and usually much more visually interesting) generative design space to work with.
No comments:
Post a Comment