Vectorized resynthesis of the previous post smart tesselate image (first one). The resynthesis is based on generating a CLIP image embedding and then 'pushing' it with a text embedding and then using the pushed image embedding for direct image to image resynthesis. If you just tried to work with that text embedding in conventional direct text to image synthesis you always get things that look like the representative image shown below. You will never get any of the new visually interesting things you get when you work with manipulating the image embeddings in a resynthesis approach. By visually interesting i mean things that aren't representative stereotypes of the images used in the original model training set.
No comments:
Post a Comment