Gen AI tests – It’s amazing how fast things change
This is an example of how fast things move. As I was doing this test, a new tool appeared that would dramatically make this process easier. By the time I downloaded it, there was an even newer tool that simplifies the process even more radically.
This test is a combination of creating a more robust LORA of myself and then swap out the background whilst keeping the main look of ‘me’ the same. Note the lighting changes from the red lights etc that appear on the face in a natural way.
It’s always good to use these these tools as quick sketches, knowing that the entire workflow will be different the next time !
This test was to explore the idea of retaining the visual consistency of on-screen talent that is potentially changed during ‘Style Transfer’ or other V2V methods, by creating a LORA of the same actor. This would be added during the generative time. This method would hopefully gain any lighting cues that only exist from the generative model.
So, First I made a bit more robust ‘LORA’ with approximate 80 images.
Rather than film a simple static scene, I opted to create a scene that might be more typical in production – something with movement and showed various scales of the person.
Although low res, This test retained the exact camera move and it does keep consistency on the face, note the body clothing has been changed completely, leaving the face with a more consistent look to the original.
Instead of relying on the Ai model to create the correct camera movements. I shot them on my phone, and used a combination of tools to swap the background and then me out of the shot.
I then did another tests that retained the same camera movement, but kept a more realistic submarine environment.
The last tests gained the benefit of using the newer models that have recently been released. I reused the LORA here, and used Veo2. While Veo3 is amazing, Veo2 has a better toolset for now. I made 2 versions, one only retaining the head, and the other a complete body swap.
Please ignore my sudden weight gain in this shot.
Conclusion
Although the test was hampered by lack of V-Ram, and these are simple tests, it’s clear that LORAs are quite transportable, and they do help to re-inforce the consistency of a live action character , and also gain lighting cues during the generative process..