I wanted to extend that scene that was created in Luma (https://sites.duke.edu/ddmc/2026/02/05/robosteve-more-testing-with-luma-ai/)
I used ChatGPT to take a frame of the Luma video and and have it create a “model sheet”. This is a common pre-visualization tool that allows a human artist to see a character in multiple angles. I wanted to use “Flow” (GoogleLabs) to repeat the same character so I wanted to give it the most information since I would be showing a full sized character (or at least more of the character) and I didn’t want to make Flow visualize the character AND create a video. That generally doesn’t give the best results. ChatGPT did not do a great job. Not only did it not follow the prompt, but it didn’t give the detail in the arm.
I swapped over to Gemini to give that a try. To be fair, I did extend the prompt to make sure it took attention to detail.
This did a good job. It got the arm detail but didn’t get the helmet detail correct. That is my fault. I should have given Gemini the reference of the side of the helmet from Luma as well. Again, this is just a one pass proof of concept. I’m only working part time you know.
I then took the model sheet and that frame from the Luma video and put it into Google Flow (Veo 3.1).
Not perfect, but way better than I’ve seen. It did a great job recreating the set. Better than I’ve seen from just one image. I tried to resize the Luma video to match the last frame, but having the different helmet styles made it more distracting. The placement of the guitar and the placement of the planet were not matching.
When I have time, I’ll try to create more samples for Gemini to create just the environment and more angles of the character, if it exists. If it didn’t exist in Luma, then Flow wouldn’t know that it wasn’t right.
Here are the two videos.
Flow also added some random text in there… par for the course.
