Automated Creativity for Brand Campaigns

Where this sits, in the four layers of DCO

The table below breaks DCO down into four layers, to better scope this project. Each row states that layer's main task, and explains whether it's compatible with an independent project, detachable from the surrounding infrastructure. The first row meets these requirements, since it is a self-contained module which, by adding an I/O interface, can be slotted into a DSP wrapper. The ones that follow, for different reasons that boil down to a few recurring constraints — integration, budget, fixed and elaborate communication protocols, data, and so on — make compatibility practically impossible.⁴

Layer	What it does	Addressed here?
1 · Variant production	Creating the creative material the system will then distribute	YesMulti-shot sequences with a consistent cast, automatic shot-by-shot checking, export into formats and feed rows
2 · Assembly and decision	Rules, templates, feeds, live signals: combinations monitoring	NoIt would require the feed integrations and ad-serving contracts the brand already holds; rebuilding them in parallel would duplicate infrastructure without improving the creative by one inch
3 · Delivery	Buying the space and putting the ad on air	NoThis is delivery infrastructure and certification, not a creative problem
4 · Measurement and optimisation	Reading the results and moving budget onto the variants that pay	NoIt would need field data and traffic volumes a production supplier does not have; simulating them would cost more than they are worth and would produce numbers no planner would take at face value

In short, the only flexible point in a "bottom-up" approach is the first layer — the one responsible for the creative scope of the content that carries the message and reaches the consumer directly. The layers above it belong to DCO's overarching management and structural layer, which involves the agency's strategic choices on budget allocation and its strategic objectives across different media.

Who it's for

Who is this for? Performance-marketing teams and the agencies that work for them — anyone who needs the same characters and product to stay consistent across dozens of ad variants, without redrawing them one by one for new settings or new character interactions. It sits upstream of the dynamic-creative platforms already in use — Flashtalking, Celtra, Criteo DCO+, DV360/Studio — not in competition with them. It fits into a pre-production stage that feeds richer, more coherent variants directly into whichever of those platforms is already running. We're not talking about the big brands running national campaigns through the major agency hubs. The realistic entry points are narrower.
The first kind of client is the mid-market digital agency. They're not looking for a platform to install in place of their own stack, but for a supplier to lean on for the creative side — much as they lean today on a freelance photographer or an outside illustration studio. A concrete example: a regional agency running paid social for a five-location restaurant chain, not a large multinational account. The second is brand marketing teams with a recurring mascot or character, where the character's coherence across a story matters more than simply swapping the product inside a template. This is the typical case for food & beverage, wine, or regional FMCG (fast-moving consumer goods) brands: think of a pasta maker with a long-standing character on the label, who needs new narrative variants, not just new products. The third is pilot projects, run alongside the DCO a brand or agency already uses, without replacing it. Example: a brand already on Flashtalking wanting to trial a richer pool of creative variants for a single seasonal campaign (Christmas, the grape harvest), before deciding whether to extend it across the whole year. In this case the engine doesn't take the place of the existing DCO, it feeds it with more material to choose from.

The benefit

The tool works as a chain of five stages:

• Input — what an agency has already decided: brand and slogans, a locked cast of characters, a setting, a story.
• Engine — the story is staged shot by shot, for each slogan.
• QA — an automatic review model checks every image against the cast, and corrects or discards the ones that drift.
• Formats — every approved shot is recomposed into the formats that will be needed, keeping the cast in frame and the area clear where the template will write headline, price and CTA.
• Export — the pool leaves as a structured Manifest (one data file, one row per shot, format and slogan, each with a unique ID, variant labels and QA status), ready to enter the feed of the dynamic-creative tool the brand or agency already runs.

The last two stages — formats and export — are covered in detail in The Engine: generation, control, export (Chapters 06–07); DCO and the Agency Market covers how the hand-off to the dynamic-creative tool works.

What the tool adds is variety, not a new targeting method: the dynamic-creative platform has more coherent material at its disposal to choose from. One industry report puts DCO campaigns at a 32% higher CTR and a 56% lower cost-per-click than a single static creative.

What it does not do

Finally, the tool we're proposing, in its current prototype form, has two clear limits. The first concerns live-action footage of real people: the engine generates and keeps illustrated characters consistent, not filmed actors. It's not built for campaigns where the starting material is video of real people — that needs something else. The second concerns absolute legal certainty. The correction loop's automatic review step sharply reduces how many shots need a human review, but it doesn't remove that need entirely: there's always a small chance that a flawed image slips past the check unnoticed by a supervisor.³
A fuller map of the landscape introduced earlier on this page, in "Who it's for" — the specific agency tier this targets, and the honest pros and cons of reaching the market through that channel — is discussed in The Agency Channel in Italy.

Pros and cons

What works well

• Character identity stays consistent across dozens of narratively sequential staged shots — it isn't a single background template, but characters, faces, clothing, setting and lighting that all adapt to one another
• Produces a richer variant pool without touching the brand's or agency's existing DCO/DSP contracts, feed integrations or delivery infrastructure — it's additive, not a replacement sell
• A closed-loop QA step (generate → judge → correct) catches off-model frames automatically, instead of relying on a human to spot every one
• Concretely: once the key characters are locked, each additional or modified staged shot costs one more engine run, not a new week of production — the marginal cost of the twentieth variant is very close to that of a second variant.
• The bigger volume advantage isn't one campaign needing a hundred variants — few agencies do. It's the engine run as a shared service: the same formats, stages and QA loop apply regardless of brand or brief, so one centralised system can run variant batches for many different agencies and inputs in parallel, each paying only for its own run, not for rebuilding the pipeline
• The pool isn't only inventory ready to ship: it also gives the team that makes the creative call something to compare side by side before deciding — a handful of staged options to discuss, not one AI-rendered guess to approve or reject

---

Current limitations

• Consistency under motion, video or live-action is still the field's frontier — here, we don't yet offer an improvement over traditional live-action or today's AI-video tools. What the engine produces is a sequence of still frames, coherent in narrative progression for the storyboard.
• A remaining gap — not yet addressed, and also barely documented elsewhere — is the direct hand-off between the storyboard pipeline and ad-serving: a direct, structured link between the storyboard's output (the frames, the metadata, the variants) and a programmatic ad-serving/DCO system — that is, the platform that actually publishes and distributes the ads across channels (trafficking, the real-time dynamic swap of format/slogan toward the delivery environments: Meta, Google, various DSPs).

---

The limit is turning the sequence of images we used to generate the storyboards into continuous video, with camera movements and transitions. That's a different problem, still largely unsolved in a fully automatic way, without human supervision. Video originally generated by AI, with consistent characters and settings, still needs costly, specialised human supervision today — continuity, VFX review, filters, colour grading, audio and image-smoothness correction, and so on. The engine we're proposing simply sits upstream of that stage, automating the pre-production storyboard that used to need an illustrator; the move to finished video remains, with or without our contribution, a separate later step.

Either way, our still-frame automation doesn't lose relevance as AI video matures — if anything, it becomes even more necessary, since Kling, Veo and Higgsfield technically require a reference frame/start-end frame to achieve consistency, given that character “drift” is mitigated today precisely by that reference frame. The AI-native sector also explicitly treats the storyboard stage as an approval gate before moving on to the real costs of the finished video.

1. Google's tool is horizontal by design — it has to work the same way for a shoe brand, an airline and a bank. That pushes it toward the lowest common denominator: swap an image, swap a price, swap a CTA. Character-consistent, multi-shot narrative generation — the same mascot staying recognisable across a dozen different scenes — is a narrow, vertical problem that a horizontal platform has no incentive to solve well, because most of its advertisers don't need it. ↩
2. The problem with using AI, which is indeed automated, but at the level of the single frame, and almost never at the level of a continuous sequence of shots that stay consistent within a story. The difference between the two levels comes from the fact that a single generation forgets the one before it, and keeping a memory of context requires a layer above a simple prompt, however well engineered. That higher layer comes from more complex architectures, built from chains of agents calling APIs or from node-based structures like ComfyUI's. Whereas the off-the-shelf tools that promise, in one shot with a few parameters and a prompt, consistent characters and settings across at least 10 shots, are often poor and need substantial manual touch-ups. ↩
3. For a pharmaceutical brand, or a campaign in a heavily regulated sector, where even a single out-of-compliance frame is an unacceptable legal risk, that residual margin can be intolerable: there, every single shot needs a human review, not just an automatic spot-check. ↩
4. What the pool, by itself, does not do: (1) it does not bid — someone still has to decide what a single impression is worth; (2) it does not choose the audience — we declare what a scene is, never who should see it; (3) it does not react in real time — a new weather or price scene isn't painted on the spot, the platform switches to, or overlays onto, a scene already sitting in the pool. ↩