BREAKING IT DOWN:
1. Creative Development:
I set out with the idea to do a test TVC for a premium fragrance brand I made up, to see if I could land an elevated aesthetic and a dynamic visual story using just AI tools. I wrote the script cold, then fed it into Claude to break out into scene breakdowns and a shotlist, which it does easily with a fair amount of editing. That got us to the prompt list – the next step on the way to generating the reference frames for video. These frames are some of the base elements I evolved through an exhaustive process of design – character, sets, props and products, wardrobe, hair and make-up, lighting, photography and cinematography. It was a lot of work, but a very creative process, so I really enjoyed it. Having these locked is key to consistency through the video generations as the video models rely heavily on reference images. Side note: I was glad to be able to use an “unlimited” image model (I used Nano Banana Pro in Higgsfield) for this part of the process as it needed plenty of generations.
2. AI Model Selection
This image of the woman below is my original cast and styling shot (which I love) that I’d hoped would carry more fully from Midjourney into Nano Banana Pro. Compared to the shots above, you can see the aesthetic slipped across models. But I think you have to accept the limitations at present – Midjourney can’t hold character and environment as consistently as NBP, while video models in Higgsfield like Kling and Cinema Studio do a better job. For the VFX sequence, Kling was good, while I came to rely on Cinema Studio as I went hard on the video generating through the dance sequences, as it was much better for continuity. Both lost a bit aesthetically from the original look, though, and the character consistency and continuity was really hard to maintain between models, as shown in the VFX frames below, which I had to use in the edit because I just couldn’t get an exact match. I imagine this kind of thing will become more accurate and easier to navigate in time.
3. VFX:
The concept here is very VFX heavy, which was a deliberate choice to test current capability. And the results are pretty good. Compared to what a high-end VFX studio could create, it would fall short on execution, but there would be no comparison on cost. The transition from the real to the imaginary world is really fun, and close to what I had briefed, but not perfect. I think the difference is what a good VFX artist would bring to the table, with tools that offer a much higher level of flexibility. That’s if we’re talking broadcast-level final delivery – as a pre-visualisation, I think it’s pretty killer.
Overall, the dance sequence feels like a success, despite some character and wardrobe inconsistency – a result of generating video across models and also just the quirky nature of gen-AI video. Nothing some quick cuts couldn’t make work for a pre-vis, if not for a high-end TVC. The level of finish on the 3D bottle is close, but not quite there compared to high-end VFX. And the bottle kept randomly changing size! Also, I would have liked more camera angles to play with, but the models struggle with logic, consistency and continuity when you, say, try for an overhead angle. Even a simple close-up is un-necessarily tricky, the easiest option being outputting at 4K to give latitude for reframes. Which is the kind of thing I would do if shooting with cameras, but with each generation burning real $$, the cost can add up when so many aren’t quite right.
4. MUSIC:
Having spent too much of my lifetime trawling through library music catalogues, this process was a game-changer. I prompted Claude with “an old-school tango track with some varied movements and rhythm to allow for dynamic editing. Authentic instrumentation and recording style.” Claude interpreted that as “A tango track with various movements featuring rich orchestration, smooth bandoneon melodies, and elegant rhythmic patterns that evoke the classic Argentine tango atmosphere of the mid-20th century.” And this is the result out of ElevenLabs. I wouldn’t put it up against an original composition with real people and instruments, but it’s perfectly fine for a test project – and super-easy.
If you want an honest, hype-free appraisal of how AI video could work for you, I’d love to chat. Email here.