The Server Never Touches the Video: IKEA Mexico's First AI-Powered Campaign

Overview

The stakes make it worth telling. IKEA is still relatively new to the Mexican market, and Ofrenda was its first campaign that both ran on its own website and handed each visitor a generated, personalized asset: a family photo turned into an animated Día de Muertos keepsake, seconds after upload. There is no template for that at IKEA Mexico yet. We were writing one.

The whole edit lives inside a URL: a Cloudinary transformation string that describes the composite, and nothing else. The CDN does all the video processing, so none of it ever touches our servers, which is the whole point, because this was a campaign built for massive traffic. This is the story of how that string got written, and why I now believe the hard part of AI video isn't generating pixels. It's everything around them.

Ofrenda IKEA title card: a Kallax shelf dressed as a Día de Muertos altar with candles, marigolds and sugar skulls on IKEA yellow, with the tagline "Un espacio para los recuerdos" — Campaign collage: Ofrenda IKEA, un espacio para los recuerdos

The ofrenda frame template: a Día de Muertos border of sugar skulls, candles and roses in IKEA blue, with an empty center where the generated portrait plays — Campaign collage: Ofrenda IKEA, un espacio para los recuerdos

And the deliverable itself, playing on loop: one photo, animated, composited into the ofrenda frame.

Why It Matters

Calling a video model is no longer a skill. Anyone with an API key can generate five seconds of moving pixels in an afternoon. What separates a demo from a product is the last 10%: the brand compliance, the aspect ratios, the cost caps, the iteration speed. All the unglamorous constraints that a campaign lives or dies by.

The URL-as-pipeline idea solved the biggest one. In a traditional motion-graphics workflow, every change means a re-export: open the timeline, nudge the overlay, render, upload, store the new artifact. Multiply that by every client revision and every A/B variation. It's v7_FINAL_final.mp4 all the way down.

With transformation URLs, the edit is a pure function. The generated video is stored exactly once. Want the overlay 8% smaller? Change overlay_width from 235 to 218 and you have a new URL. New frame, new font, new text position — different arguments, same asset, zero re-renders. A parameter change is a deploy, and every composite is processed at the CDN's edge, on their compute, not ours. For a massive campaign, that difference is the architecture: no transcoding fleet to size, scale, or pay for, no matter how many visitors show up.

Tech Stack

FastAPI on Vercel serverless Python: the conductor
Gemini 2.0 Flash via a pydantic-ai agent for validation
nano-banana (Replicate) to retouch the portraits
wan-2.2-i2v-fast (Replicate) to generate the motion
Cloudinary transformations as the compositor
Supabase Postgres + Storage for the memory

What It Does

The pipeline is a relay race, and each model runs one leg.

Gemini is the bouncer

Every upload hits a Gemini agent with a strict-JSON contract: is_valid_public, subject_type, is_portrait. Humans and pets only, no nulls, no prose. I wrote the prompt the way everyone writes those prompts: "Respond with ONLY a single JSON object and NOTHING else."

The model did not always comply.

So underneath the prompt sits _safe_json_parse, a function whose entire job is to scrape the first {...} block out of whatever the model actually said. Prompting is negotiation; parsing is law. I learned this building a chatbot that could never break character, and the lesson transfers perfectly: the prompt asks nicely, the code enforces.

nano-banana is the retoucher

Valid portraits get conditioned by nano-banana before animation. The interesting part is one line in the prompt: "crop/zoom only; do NOT alter identity or features."

When the input is a photo of someone's late grandmother, that stops being a prompt-engineering flourish and becomes the ethics policy of the entire product. The model is allowed to reframe the face. It is not allowed to improve it. Identity preservation as an engineering constraint, not a disclaimer buried in terms of service.

wan-2.2 is the actor

The retouched portrait goes to wan-2.2-i2v-fast: 81 frames, 16 fps, 480p, with the prompt "he dances in peace, in slow motion." Every generated video in this campaign has that same gentle motion, because the shelf it lives in is an ofrenda, and the movement has to feel like a memory, not a deepfake.

Cloudinary is the editor

The generated clip uploads to Cloudinary, and then the edit happens, as text. One transformation chain says: overlay this video onto the animated Kallax frame, force 2:3 aspect ratio, fill from center, round the corners 13 pixels, stamp the family member's name in Montserrat at 30pt in #0058A3, pick a safe codec.

That aspect ratio was a war. The overlay kept coming back distorted: faces stretched, shoulders wrong. Every prompt tweak on the generation side did nothing, because the distortion was happening downstream, in the composition. c_scale resized without caring about shape. The fix was two parameters: ar_2:3 plus c_fill with center gravity. Crop the overflow, keep the proportions. There's a commit in the history that is literally just fix(cloudinary): enforce 2:3 AR using ar_2:3 + c_fill, and it's one of my favorites, because it fixed a customer's grandmother.

The chain we ship, parameter by parameter:

l_video, the overlay. The generated clip is referenced by its Cloudinary public ID, so the composite always reads the one stored asset. Nothing is duplicated, nothing is re-encoded.
ar_2:3, the portrait prison. The Kallax slot is portrait; the generated clip is not. ar_2:3 declares the target shape.
c_fill with center gravity, the crop. This is the one that ended the war. c_scale resizes into the target shape and stretches whatever doesn't match (that's where the stretched faces came from). c_fill fills the frame and crops the overflow from the center: proportions preserved, faces intact.
r_13, the rounded corners. Cosmetic, but it's the difference between "AI video" and "video that belongs in the shelf."
l_text:Montserrat_30: with co_rgb:0058A3, the name plate. The family member's name in Montserrat at 30pt in IKEA blue, positioned under the frame.
f_mp4, the codec. Fixed explicitly. Never let the CDN guess the delivery codec on a paid campaign.

The aspect-drift debug recipe. When a composite comes back distorted, suspect layers in this order:

The composition, first, always. Aspect drift is almost never a generation problem: the model drew a fine face; the CDN stretched it. Check the transformation chain before you touch the prompt.
The target AR vs. the source AR. Print the natural dimensions of both the base frame and the overlay. If they disagree and anything in the chain uses c_scale, that's your bug.
Reproduce with a still image. Swap the video overlay for a static picture and run the same chain. If the still distorts, the bug is in the parameters; if it doesn't, the bug is in the video asset itself.

The exact transformation string

The full compose_video_with_center_overlay chain we ship, parameter by parameter, plus the debug recipe I use for aspect-ratio drift (which layer to suspect first, and why).

Want to discuss the system directly?Book a 1:1 call

Supabase is the memory

Users, consent (accept_mkt), and every generated video land in Supabase, with row-level security so nobody reads somebody else's grandmother. And when storage fails (serverless storage occasionally does), the endpoint streams the raw MP4 straight back to the customer anyway. Never send someone away empty-handed from a memorial.

Impact / Lessons

What stuck with me:

The edit lives in the URL, not the file. I stopped thinking of compositions as artifacts. They're expressions now: arguments against stored assets, evaluated lazily by a CDN.
Rate limits are product design. A low requests-per-minute limit is as much a cost cap as it is abuse prevention. The invoice math for an AI activation gets decided in the router decorator, not in the cloud bill review.

Every request pays three models, and they are not equal partners:

Gemini validator. The cheapest call in the chain, and the only one with a free failure mode: if it rejects an upload, the request costs almost nothing. The bouncer is also the refund.
nano-banana retoucher. A single image generation, an order of magnitude above the validator but still rounding error next to the video call.
wan-2.2-i2v-fast. The bill. Per-second video generation dwarfs everything else in the pipeline combined.

Which dictates the limiter strategy: cap requests where the cost is. The validation endpoint gets the generous limit (it's cheap, and friction there kills conversion); the generation endpoint gets the strict one, because that's where the invoice lives. Cost-capping an AI activation is really cost-capping one call, the one that generates pixels over time, and letting everywhere else breathe.

The rate-limit math

How to cost-cap an AI activation before the invoice caps it for you: the per-request cost breakdown across the three model calls, and how I chose the limiter values.

Want to discuss the system directly?Book a 1:1 call

Brand guidelines will out-vote your taste. I built a whole endpoint for uploading custom fonts, tested a romantic script face (Parisienne) for the name overlay… and one commit later, fix(overlay): swap Parisienne with Montserrat replaced it. IKEA blue is #0058A3 or it isn't IKEA. I stopped grieving quickly.
The moat is the last 10%. Generating the video took an afternoon. Making it correct took everything else: validated, identity-safe, aspect-true, brand-exact, capped, and resilient to Supabase having a bad day.

The most useful shift in my own head: I used to ask which model should generate this? Now I ask where exactly does the model's job end and the code's begin? Draw that line early, enforce it server-side, and the pipeline survives contact with real customers.

Built with FastAPI, Replicate (nano-banana, wan-2.2-i2v-fast), Gemini 2.0 Flash, Cloudinary, and Supabase.