I boosted the story line prompt strength relative to the style prompt for these runs. Both recursively feedback the last frame (affine warped) to initialize the latent space of the U-Net. Above is using a .5 'contrast', and below is using a 2.0 'contrast setting, which is just a multiplier for the previous frame tensor (which then has some noise added to it before being used to initialize the latent space).I also just realized that everything i've been running with recursive feedback so far has been incrementing the seed each frame. Ouch, not what i was assuming was happening. So we need to run this particular experiment again with a true fixed seed.
A depository for John Dalton's personal artwork. Studio Artist, MSG, procedural art, WMF, digital painting, image processing, human vision, digital art, slit scan, photo mosaic, artistic software, video effects, computer painting, fractals, generative drawing, paint animation, halftoning, video effects, photo manipulation, modular visual synthesis, auto-rotoscoping, directed evolution, computational creativity, artificial intelligence, generative ai, style transfer, latent diffusion
No comments:
Post a Comment