Everyone asks which model. Wrong question.

Models are the cheapest part of the stack

We run eleven frontier video models plus the current state of the art in image and audio, and we swap them the week something better ships. That is deliberately unremarkable. Model quality is a moving target every vendor buys at the same price, and a creative team does not want a model. They want an ad that clears legal, matches the guideline deck, works in Arabic and fits nine placements.

The marketing harness: brand lock, layout grammar, type discipline, safe zones, format rules, legibility checks. Same models everyone has, constrained by rules a creative agency would recognise. The harness is why output looks briefed rather than prompted.

A canvas of nodes, each shaped like a real agency task

Creative Studio is node-based. You wire a graph, and each node does one job somebody in a creative department already does by hand. We did not invent these steps. We sat with creative agencies and shipped the steps they were already running across spreadsheets, Slack threads and eight open tabs.

Brand Pack takes one approved master and returns every ratio without re-cropping the logo. Localise mirrors a layout into Arabic and retypesets it rather than flipping the canvas. Composer handles compositing in layers you can still edit. Nothing here is a generic prompt box with a label on it.

Brand: every downstream node inherits the lock. Drift is not a setting you remember to switch on.
Brand Pack: one approved concept, every placement ratio, logo placed rather than regenerated each time.
Localise: Arabic and other right-to-left markets get mirrored layouts and correct typesetting, not a flip.
Composer: layers stay separable after generation, so the last 10% is an edit and not a re-roll.

One locked master. Thousands of variants.

This is the part agencies actually pay for. The bottleneck was never making the ad. It was making the approved ad again, four hundred times, once per SKU and once per placement, without one of them drifting.

So Bulk Generate takes a single layered master, the same PSD a designer already signed off, and treats it as a contract. The offer block, the campaign dates, the payment badges, the helpline, the safe zones: frozen. Only the slots you mark as variable move. Point it at a catalogue and it returns the whole set.

What the contract holds

Locked: offer block. Sale lockup, the 30% figure, foil treatment and drop shadow. Pixel identical across the set.
Locked: campaign dates. 24/06/2026 to 30/06/2026, written once. Change it once and all 530 update.
Locked: trust furniture. Shop with Confidence badge, helpline, Buy Now and PayLater marks, 24 hour delivery stamp.
Variable: product scene. The only region that moves. Bedroom, sofa, wardrobe, recliner, dining, one per SKU.
Variable: ratio and crop. 9:16 for social, 6:5 for programmatic, and the rest. Recomposed rather than cropped.
Variable: language. The same contract re-typeset for Arabic, with the layout mirrored.

Look at any two of the HomesRus ads. The red block, the foil numerals, the date band and the badges do not move by a pixel. That consistency is not discipline from whoever was on shift. It is the contract holding.

1 layered master signed off by the client
530+ finished ads out of it, in under four weeks
9:16 and 6:5, social and programmatic from the same contract
Weeks to minutes: production time per placement set

Not mockups. Paid media, in Doha.

The harness is easy to claim and hard to fake, because out-of-home has no forgiving crop. A board is four storeys up in daylight, at a fixed ratio, with a client's legal team already through it. This is campaign creative produced on ElarisLabs and bought as paid digital OOH in Qatar.

For McDonald's in Doha, look at the first and the third board. Same campaign, same shoot, same talent. One board carries an English headline ("Feels like 0°C") over a cold-weather palette. The other carries the Arabic line with the layout mirrored, the numerals retypeset and the logo moved to the correct corner, keyed to a 42° day. That is the Localise node doing the thing a canvas flip cannot do.

For Qatar Development Bank, with ELAN Media: four SME brands under the QDB programme (Elite Food Industry, Fresh Meat Factory, QLife Pharma and Crystal Food Oil), each with its own lockup and its own product, all produced through the same graph. One workflow, four brand contracts.

10.4M impressions on the live McDonald's Qatar OOH run
9 live screens carrying the campaign
EN / AR bilingual sets from one approved concept
Paid: bought media, not a portfolio render

A number before the media spend, not after

Every creative the canvas produces comes back with a predicted performance score against the brand it was made for. Not a generic aesthetic rating. Scored on the things that decide whether an ad works: whether the message survives a thumb-scroll, whether the CTA is findable, whether the brand is identifiable in the first half second.

The point is triage. When a run returns 530 finished ads, somebody has to decide which eight go live. That decision used to be a meeting.

Predicted score: 84. Brand recall in 0.5s: 91. Message clarity: 88. CTA prominence: 76. Scroll-stop likelihood: 81.
It ranks the set. 530 variants sorted by predicted performance, so the shortlist is built before the review call rather than during it.
It closes the loop. Once the campaign runs, real results come back and get compared against what was predicted. The score gets better because your own campaigns taught it.

Pull your team onto the node, not into a thread

Creative review is where most production time actually goes, and it almost never happens where the work lives. Feedback arrives as a screenshot in a group chat, and somebody translates it back into a change. So we put review inside the canvas.

Call anyone on your team onto a graph. They comment on a specific node, a specific layer, a specific variant. Everyone sees the change as it happens. The comment is attached to the thing it is about, which means nobody has to ask which version.

The graph you built is the template

The interesting thing about a node graph is that it is a record of how the work was done. So once a multi-step ad is built and approved, save the whole graph. Next month the same campaign shape runs against a new brief, a new catalogue or a new market, and the process does not get rebuilt from memory.

This is the difference between a tool that makes one asset fast and a tool that makes a team's process repeatable. The HomesRus run is not eight clever images. It is one saved workflow that can be pointed at next season's catalogue tomorrow.

From the same node, live

Approved creative pushes straight to your ad accounts (Meta, TikTok, Snapchat, LinkedIn and more), or onto the social scheduler. The calendar optimises placement timing per channel and writes the captions in the brand's voice, so the handoff from creative to distribution is not a download folder and a separate login. Captions are generated per channel, not copy-pasted across all of them.

Published is the middle of the job

Once it is live, results come back into the platform and get analysed against the creative that produced them. What worked, what did not, and which of the 530 variants actually earned their placement.

Then the part nobody builds for. Creative fatigues. Performance on a winning ad decays on a curve that is predictable if you are watching for it and invisible if you are not. We flag decay on live campaigns and point back at the node that made the asset, so the refresh is a re-run rather than a new brief.

Your brand now has an audience of machines

A growing share of buying decisions starts inside an answer engine rather than a search results page. If a model does not know your brand, or confuses it with someone else, that is a distribution problem with no ad slot to buy.

So the platform checks it. We run your brand and category questions across the major answer engines and report whether you get cited, who gets cited instead, and which sources the engines are actually pulling from. It sits alongside the campaign reporting because it is the same question in a new channel: did anyone see us.

Everything that happens is recorded against the brand

Every generation, every approved asset, every comment, every score, every published result, every decay flag. It all writes back to one brand record. That record is what makes the next campaign cheaper than the last one.

This is why the whole lifecycle has to live in one place. A scoring model that never sees real results stays generic. A localisation step that does not know which Arabic headline got approved last time relearns it every time. Split the lifecycle across five tools and the memory is the thing you lose.

What model access gives you, and what a platform does

Generate. A model wrapper: a prompt box on a frontier endpoint. ElarisLabs: task-shaped nodes built with creative agencies.
Stay on brand. A model wrapper: you paste the hex into the prompt and hope. ElarisLabs: brand lock inherited by every downstream node.
Scale a catalogue. A model wrapper: re-prompt per SKU, the offer block drifts. ElarisLabs: one locked master, offer block frozen by contract.
Every placement. A model wrapper: re-prompt per ratio, logo redrawn each time. ElarisLabs: one master, every ratio, logo placed not generated.
Arabic and RTL. A model wrapper: flip the canvas, break the typesetting. ElarisLabs: mirrored layouts, retypeset, safe zones held.
Decide what ships. A model wrapper: a review meeting. ElarisLabs: predicted score on every variant, pre-spend.
Review. A model wrapper: screenshots in a group chat. ElarisLabs: comments pinned to the node, live for the team.
Repeat the process. A model wrapper: rebuild it from memory. ElarisLabs: save the graph, re-run against a new brief.
Distribute. A model wrapper: download folder, separate login. ElarisLabs: push to ad accounts, scheduler, optimised calendar.
Learn from results. A model wrapper: nothing comes back. ElarisLabs: results analysed against the creative that made them.
Keep it working. A model wrapper: notice the drop next quarter. ElarisLabs: decay flagged on live campaigns, refresh from the graph.
Get found by AI. A model wrapper: not a feature. ElarisLabs: brand AI visibility tracked across answer engines.

One canvas for everything an ad needs

A brand profile, the generation, the harness that makes it look briefed, the locked master that fans out into a catalogue, the score that says what to ship, the team that signs it off, the schedule that puts it live, the results that come back, the decay flag that says refresh, and the memory that keeps all of it. In one tool, because the loop only compounds when nothing leaves the building.

That is what we are. Not the model. The rest of it.

Brand → Create → Score → Review → Publish → Measure → Refresh → Brand memory