woman dancing -part 3

 

One shot resynthesis of the first frame of the original source video used for this series of posts (shown below).  Same static text prompt used in the other posts is used for the dual-guide resynthesis divergence as the series progresses.  This is using the new generalized latent diffusion framework, so it's not stable diffusion anymore.

I tried 2 runs processing the first 100 frames of the original source movie (see part 1) with a fixed seed resynthesis, shown below.

These 2 video processing runs are both pretty disappointing from the standpoint of video processing or generative animation.  Need to run some more experiments, but for now i give the generalized framework an A+ for single image resynthesis, but it probably needs some additional work and/or restructuring to be really useful for this kind of approach to animation.