What surrounds the model

The question I get most often: "What do you make it in?"

I'll answer. But a warning first: the tool matters less than you'd think. The model is the easy part.

The tools we use today

Gemini for images. Key frames, scenes, the first frames that motion is later built from.

Seedance for image-to-video. It takes a first frame and can hold the last one too, so a shot lands where it should and connects to the next.

Omni for talking characters. When a mascot or a character has to speak to camera.

ElevenLabs for voice. Voiceover in the campaign's language when there's no time or reason to book a studio.

Remotion and HyperFrames for graphics. Captions, typography, motion, anything that has to be frame-accurate.

Blender for a rough 3D scene before anything is generated. Where the camera stands, where the character is, where the light comes from. Trying things in Blender costs nothing. Only the final gets generated.

This list will change. Models come and go faster than a single campaign can be shot. That's why I don't build on it.

The model is the easy part

Anyone can use the same model today. If the result depended on the model alone, everyone would have the same video.

The difference is what surrounds it. Rules, checks, the order of steps. The technical word is harness. I call them gates: points the work has to pass through, or it doesn't move on.

In video we have four that a client will feel, even if they never see them.

1. One master, derived formats. First there is one main version. Only once it's approved do we derive the vertical, the horizontal and the wall format from it. Not five independent versions that don't match.
2. Image first, then motion. Only an approved frame gets animated. Animating a mistake is the most expensive way to discover it.
3. A list of approved shots. A generation folder holds hundreds of files, and the newest isn't always the right one. So there's a manifest: what was approved, in which version, and when. We search that, not the folder.
4. 50 frames per second. Every AI clip that goes out is converted to 50 fps. Motion stays smooth even on a wall seven metres high.

Gates that stop me too

The most important gates don't protect the client from me. They protect the work from my own mistakes. Each one came from a specific scar.

Two gates around the work. Before starting, we check whether the plan still holds. Last week's plan is a hypothesis, not the truth: a file has changed, the situation is different. When the work is done, a second gate checks the result. Only then is it finished.

A green light over a dead channel. One of our checks once stayed silent for eight days and everything looked fine. Nobody noticed it hadn't run at all. Since then another watcher makes sure the watcher actually runs. A quiet system is not a healthy system.

A gate that tests itself. A check that stops everything is as bad as one that stops nothing. As of 21 September 2026, 81 gates had their own test. It verifies two things: that the gate stops the error, and that it doesn't stop work that's fine.

On that day, 134 guard scripts and 66 specialised agents were running in total. Each one owns a single area: camera, colour, editing, data, security.

The point

A note in memory doesn't change behaviour. A gate the work has to pass through does.

That's why AI video isn't cheap. Anyone can rent the model in a few minutes. What surrounds it gets built piece by piece, again after every mistake. And that is what decides whether the model produces an accident or an image someone can stand behind.

Want to see how it would look on your project? Write to hazelnut@epicproduction.sk or call +421 904 146 414.