Rather than asking the video engine to come up with the concept from scratch, I used Gemini to first create an image which I would then ask Veo to animate.
The first image was pretty good, but it was square:
I took it into Photoshop and extended it.
The initial result was very promising:
In hindsight, I should have cut off that back conveyor belt.
For the next test, I asked ChatGPT (Duke) 3o Reasoning to create the same image and this time I gave it more explicit instructions as to the camera angle, which it almost completely followed. It did follow the aspect ratio as well so I didn’t have to extend it in Photoshop.
I liked this but it seemed too plain. so I asked it again:
This was better, but it lost the camera view.
This got the camera angle correct, but the arrow was wrong for the questions.
OK – we’ll use this one. 🙂 It’s not actually 16:9, so I did extend it in Photoshop.
I used Veo 2 first – was SO close!!
With most things AI, one thing gets fixed that was broken and then things that were working are now broken. 🙂 Here are two examples from Veo3. I used the Scene builder to export them both as one video. There’s random audio too.
I then used Gemini to try to help me to rewrite the prompt so that the three items that need to be animated can be corrected.
This is sort of interesting. I tried the new prompt into the same “project” that the old ones were in and it failed on both versions. I then started a new “project” and gave it the new prompt and it worked! This was on Veo2 also.
Still far from perfect, but I’m getting somewhat closer to a workflow….
I’ve linked a PDF of all my prompts here:
https://duke.box.com/s/fv5oor3cij7d44d7iy6udl690p727066
