For the launch of FULLCUT, an editing aid combined with an AI studio, we needed a campaign of advertising films. The stance was radical: not a single camera would be taken out. The entire campaign is generated by artificial intelligence, 100% of it: the films, the ads, the loops on the website. Here is how we did it, what we measured, and what it taught us about the craft.
Everything starts with art direction, not with the model
Before writing a single prompt, we wrote a complete visual charter. It fits in one sentence: “Blue night, hard light, soft air.” Black is never grey, it is tinted with blue. A single light source, hard, off-frame, cutting out the contours. A light mist as the only softness allowed. The assumed reference: perfume advertising. Neon, night, a single gesture, no explanation.
This charter fixes everything: a four-colour palette (a blue-tinted black, an electric blue, a cream reserved for what is human), the prohibitions (no warm light, no orange, no lens flare, no handheld camera), and the grain, fine and regular, that of a 35mm negative pushed one stop. If you notice it, it is too strong. A shot that does not respect the charter is not fixed in the edit: it is redone.
That is the first lesson: in AI video, art direction is not an option, it is the product. Without a charter, every generation goes its own way and the film looks like a moodboard. With a charter, a hundred different renders belong to the same world.


Two engines: MiniMax H3 locally, Seedance 2.5 for the films
We worked with two complementary models. MiniMax H3 runs locally, on a dedicated workstation, through ComfyUI. It is the laboratory: zero cost per attempt, fast iteration, and it is where we measured the rules that govern all our prompts. Seedance 2.5 takes over for the final films, in reference-to-video mode: you feed it precise starting images and it animates them.
What practice taught us, with numbers to back it up:
- One action per shot. Chain two actions in a prompt and the second is simply dropped. That is not a setting, it is dramaturgy: one action per shot, and you chain them in the edit.
- A logo is not written in a prompt, it is animated. Ask a model to write a brand name and it will produce one letter too many or too few, differently on every draw. The only reliable way: provide the logotype already composed as the first image, and let the model move the light around it.
- The exact spelling of named entities. Models do not correct your approximations: they render literally what you write.
- One prompt can carry six shots. We assumed a call describing six shots would deliver three. Wrong: two 30-second takes, 721 frames each, and all six shots are there, in order, at the requested durations.
- The good film is an edit of several takes. Two takes of the same prompt are not equal shot for shot: one has the better timeline, the other the better empty chair. Choosing, assembling, matching cuts: exactly the work we do every day in post-production.
Prompts written like scripts: the JSON
For the campaign’s main film, the prompt is not a paragraph: it is a structured JSON document. It declares the format (30 seconds, 24 frames per second, 16:9 master and 9:16 variant), the palette in hex codes, the fonts, the list of reference images and, shot by shot, which input serves which shot. It even carries the laws of the film: a single sound impact, the music stopping at 25 seconds, the logo strobe capped at 3 Hz for accessibility.
That level of precision changes everything: the prompt becomes a running order, versionable, re-readable, correctable, exactly like a storyboard. It is our long-standing method all over again: no shoot starts without a validated storyboard, and an AI shoot is no exception.
The result is visible in our portfolio: “The night watched everything” and the other FULLCUT films are online, in the AI Films category.
One non-negotiable rule: no client media in the models
A point of ethics, held to the same standard as everything else: the only images ever sent to a generation service are the brand assets of the project concerned. No client footage ever goes into a model, ever. The shots we need are generated for the purpose. It is the simplest and cleanest answer to the question of rights over our clients’ films. We laid out our full position in what we do and what we refuse in AI films.
Frequently asked questions
Is a 100% AI video really 100% AI?
The images are 100% generated by the models. But the art direction, the writing, the selection of takes, the edit, the grade, the sound and the on-screen text remain fully human work, done in DaVinci Resolve. AI replaces the set, not the studio.
Why generate locally with MiniMax H3?
Local means iteration without a meter: every attempt is free, and it is where we refine the prompt rules before launching the final renders on a more powerful model. It is also a confidentiality guarantee: nothing leaves the machine.
Why write prompts in JSON?
Because a film is driven like a running order, not like a sentence. The JSON declares the duration, the palette, the reference images and the role of each shot. It can be versioned, re-read and corrected: it is the storyboard of the AI shoot.
Do you use your clients’ footage to train or generate?
No, never. No client media ever goes to a generation service. The only images sent are dedicated brand creations, and the shots we need are generated specifically for that use.
Want to explore what a 100% AI film, or an AI-augmented one, can do for your brand? Let’s talk: reply within 24 hours, with a frank opinion on what AI can and cannot yet do for your project.