From 3D to AI: same dino, new approach

There were several reasons for this. First, turnaround time. 3D rendering and animation are labour-intensive, especially when they’re integrated into a traditional video production. That’s no criticism of the craft: that labour is exactly what delivers the control and the quality. But for a series of spin-offs where the character is already established, the time from concept to execution matters more than it does for a hero spot.

On top of that, today’s AI models are finally good enough to be useful beyond the concept phase. That meant we could develop a wide range of scenarios and present them to the client without every proposal requiring a full production cycle. That changes the conversation: you’re choosing between things you’ve seen, not between descriptions.

But the most concrete argument was interaction with the environment. Making a 3D element interact with a real environment is highly complex in traditional production. For the TV commercial, for instance, we had to source a real car wreck and dent the roof by hand, all for a single scene. And in that scene, you don’t even see the dino denting the roof, only the aftermath: that was as far as we could go within the time and budget available. In 3D production, every shadow and every reflection of the dino takes work to create and to integrate properly with the footage. Thanks to their holistic approach, AI video models generate that interaction in one go.

We worked in four steps, each with a client sign-off.

1. A character sheet of its own

   The dino in the spin-offs is inspired by the one in the TV commercial but differs in the details. Think skin texture, eyes and teeth. We started by creating a character sheet showing the dino from various camera angles and with different expressions, plus close-ups of its feet and skin details. Once the client had approved it, we moved on.

2. A style for the environment

   The world around the dino is stylised: ACM purple, with simple geometric shapes as background elements. The objects the dino actually interacts with are more detailed and realistic than the rest of the environment, so the eye goes where the action is. We also chose not to show any people. That means no uncanny valley, that unsettling feeling when something looks almost human but not quite, and no photorealism to magnify the weaknesses of the AI models.

3. One still per scenario

   For each scenario, we created a still and refined it with the client until we had a final image. That image was then upscaled to high resolution and served as the reference for video generation. So, the still wasn’t an intermediate step we could skip. It was the anchor for the moving version.

4. Post-production in After Effects

   The generated videos aren’t the end product. In After Effects, we handled the colour grading, scene extensions and timing adjustments, and added motion design elements.

This approach certainly isn’t without its hurdles. Take resolution. Even the best models still can’t generate razor-sharp 4K with rich detail. In fact, some of today’s strongest models top out at 720p. Upscaling isn’t an option; it’s a mandatory step.

Then there are the teething problems. Artefacts, hallucinations, temporal inconsistencies, concepts bleeding into one another: it all comes with the territory. That calls for selection and rework, and you never know in advance how much.

Another limitation is the lack of precise control. We worked with detailed text prompts and reference images, not keyframes. ‘The dino’s foot needs to be slightly higher at the five-second mark’ isn’t a tweak you simply make. It’s a new generation with a different prompt, followed by another round of selection. If you need that kind of precision, 3D is still the better fit.

What’s more, the models output video as a single flat layer. If you only want to adjust the colour grading on the dino, you first have to mask it out. That doesn’t always work, because of motion blur and artefacts. What’s a routine job in 3D becomes a chore here.

Finally, there’s the fixed format. An AI video is generated in a single aspect ratio: landscape, portrait or square. Turning it into a different format afterwards isn’t straightforward. In a fully 3D pipeline, you can simply leave everything as it is and re-render in another aspect ratio. With AI video, that’s either impossible or far more laborious: you generate a new video, and it deviates from the image you’d already approved. AI reframing tools that can change the aspect ratio of footage do exist, but for now they’re very limited and the quality of the generated footage falls short.

The 2025 TV commercial wasn’t our first outing with the dino. We created the first dino visuals for ACM Insurance back in 2024. In that video, a gigantic dino foot crushes a car. The production was done entirely in-house: we only used AI for the texture of the foot. The rest was manual animation in After Effects, prepped with a healthy dose of Photoshop.

Back then, the models weren’t ready for professional video production. Today, they’re powering an entire series of spin-offs. That’s how far we’ve come in two years. So, the limitations above aren’t an end point, but a snapshot. Because if the past two years have proven anything, it’s that this will all look very different tomorrow.