Model: Veo 3 Fast
1280x720
24fps
00:00:32
(85 shots of 192 frames, making 4 sequences with overlaying)
This small experiment takes the first frame of such a film, and extends it using image-to-video neural network Veo3. One characteristic of such software is that it is able to infinitely regenerate the same video with small variations by using different noise patterns as input for the decoding process.
Though these settings are hidden away in such proprietary models, they have taken the habit of displaying four possible outcomes of a given prompt - in this case, four videos. Each video is eight seconds long. In this tiny film, I take all four suggested outcomes, and extend them each. The resulting images are overlaid with diminished opacity so as to observe the drift that occurs between the different versions of a same image.