Archive: VQGAN Zooming

During the early days of image generation, there was a popular technique known as “zooming.” Essentially, you’d generate an image, and then apply a transform to it (translate, scale, and/or rotate) and then run again to advance the image. When applied over and over, you’d step through an animation and make a video. Additionally, for each step/frame you could also interpolate the weights of different prompts, so that you could transition from one to another.