PUMA AI Creator: Revolutionizing Fan Engagement and Digital Art with Google Cloud

Lisa Ernst · 25.01.2026 · Artificial Intelligence · 9 min

A football jersey is usually the end of a design story: a club and a brand decide, fans react, and the discussion happens after the reveal. With the PUMA AI Creator, PUMA flipped that order. Instead of “here’s the kit”, the brand asked: “Show us what you’d make.” The result wasn’t just a marketing gimmick. It was a measurable, large-scale test of what fan-led creation looks like when generative AI, cloud infrastructure, and game mechanics are stitched into a single experience.

Source: Own illustration

PUMA (one of the world’s major sports brands for footwear, apparel, and accessories) partnered with Google Cloud to bring generative AI directly into its digital ecosystem. What made this initiative stand out wasn’t that it used AI, but that it used AI at retail scale: generating images fast enough, reliably enough, and safely enough to serve a massive global audience without turning the experience into a slow, glitchy waiting room.

In practical terms, this meant moving from static “brand tells / customer listens” storytelling to an interactive loop where customers create, react, and vote — and the brand collects real engagement signals along the way.

Quick Summary: PUMA AI Creator Highlights

• Global reach: 54,000 users from 206 countries participated.
• High-volume creation: 180,000 jersey designs were generated.
• Community feedback loop: 1.7 million fan ratings were recorded via swipes/votes.
• Cloud foundation: Google Cloud Platform with Google Kubernetes Engine and Nvidia H100 GPUs.
• Speed gains: Generation time reduced from ~60 seconds to ~17 seconds for four HD images.
• Cost control: 306 Nvidia GPUs used during the live phase instead of an initial projection of 600.
• Personalization engine: Imagen 2 on Vertex AI enables region- and context-aware product imagery.
• Commercial lift: Personalized content + segmentation improved CTR and AOV; later results include measurable uplifts (details below).

PUMA AI Creator: A Fan-Led Design Revolution

The heart of the story is the PUMA AI Creator — a fan co-creation project built around Manchester City. PUMA, as the club’s official kit manufacturer, didn’t just showcase a new design direction; it invited fans to generate the direction. The goal was ambitious: create a pathway toward an AI-assisted, fan-designed jersey concept for Manchester City’s third kit (2026/27 season).

The live phase ran from December 9 to December 20, 2024. During that window, users could register and start producing concepts through a browser-based tool that combined a 3D jersey preview with text-to-image generation. The constraint was simple but important: a text prompt of up to 300 characters. That limit forces clarity — fewer “novel prompts”, more design intent.

How the experience worked (step by step)

1. Register on the platform and receive 10 design credits.
2. Write a prompt (max 300 characters) and generate jersey concepts inside a 3D browser app.
3. Save designs into a gallery and shortlist favorites.
4. Submit up to two designs as competition entries (two submission opportunities per user).
5. Vote and earn: swipe-rate other designs; every 10 votes unlock 2 additional credits, keeping participation going after the initial credits are spent.
6. Reward communities: Manchester City fans (“Cityzens”), members of the DEEPOBJECTS.ai community, and PUMA NFT holders received double credits, making the most invested groups the most active creators.

That loop (create → share → vote → earn credits → create again) is not accidental. It’s a classic engagement engine, but here it’s tied to a production-grade AI backend. The output volume shows it worked: 180,000 designs and 1.7 million ratings from 54,000 users in 206 countries — all inside a 10-day live phase.

Ivan Dashkov (Head of Emerging Marketing Tech at PUMA) framed the initiative as proof that PUMA wants to lead with technology even when there’s no established playbook. And that’s really the point: this wasn’t only about jerseys — it was a stress test for “AI + commerce + community” in the real world.

Source: sgieurope.com

Ivan Dashkov PUMA portrait. This image features a man in a PUMA jacket, smiling, against a modern hallway background. It serves as a clean and natural portrait, aligning with the "Ivan Dashkov PUMA portrait" search phrase. The image is well-lit and free of any distracting overlays or text, making it suitable for an inline blog image slot.

Under the Hood: Google Cloud and Generative AI

Cool campaigns die on boring problems: latency, scaling, concurrency, and costs. If thousands of users click “Generate” and the system chokes, the magic is gone. To make the PUMA AI Creator feel instant (or at least fast enough), PUMA needed infrastructure that could handle global spikes without collapsing into timeouts.

PUMA partnered with FTR as lead agency, Modern English for the app’s frontend/backend development, and Slalom for the AI backend and scalable cloud infrastructure. Slalom migrated a local version of the app — originally using ComfyUI + a Stable Diffusion image generation model — into a containerized deployment on Google Kubernetes Engine (GKE).

A key architectural piece was a custom queueing system to sort, buffer, and sequence user requests, designed to tolerate up to 50,000 incoming prompts. That kind of buffering is what keeps the user experience stable when demand is chaotic.

The second success factor was hardware: Nvidia H100 GPUs. In the final 15 days before launch, Slalom and a Google expert team focused on performance bottlenecks and GPU optimization — and the payoff was dramatic. Generation time dropped from roughly 60 seconds to about 17 seconds for four high-definition images.

Speed wasn’t the only win. Efficiency improved as well: during the live phase, the system ran on 306 Nvidia GPUs, instead of an initial projection of 600. That difference matters, because GPU-heavy AI workloads are where costs can spiral quickly.

Source: indiamart.com

Nvidia H100 GPU card. This image shows a side-angle view of an NVIDIA H100 PCIe card on a clean white background. It clearly displays the card's bracket, ports, and overall profile, making it an excellent representation of the physical hardware. The clean presentation and lack of overlays make it ideal for an inline blog image slot.

Why the infrastructure details matter

• Perceived creativity depends on responsiveness: if “Generate” takes too long, users abandon.
• Queues protect the experience: buffering prevents traffic spikes from breaking the system.
• GPU efficiency = cost control: halving projected GPU needs changes the economics of running similar campaigns.
• Container orchestration enables scale: GKE helps manage a distributed system under unpredictable demand.

Imagen 2 on Vertex AI: Personalization That Feels Local

The AI Creator grabbed headlines, but PUMA’s broader AI story is about personalized commerce. The campaign leveraged Imagen 2 on Vertex AI to generate dynamic, context-aware product imagery — not just “cool AI pictures”, but images that fit what a specific user in a specific place might respond to.

The concept is simple: the same product can be framed in different environments to match local culture and preference. A customer in Japan might see a lifestyle shoe pictured on the streets of Ginza, while a trail shoe appears near the foothills of Mount Fuji. That kind of localization is hard to produce manually at scale, but it becomes feasible when the pipeline is automated.

Imagen also supported PUMA’s content teams with repetitive image editing tasks — shadowing, composition, color accuracy, resolution, and product positioning. The operational impact is straightforward: less manual production time, faster rollout of campaigns, and shorter time-to-market across regions.

Broader AI Adoption and Future Plans

PUMA had already migrated key parts of its e-commerce ecosystem (including puma.com) to Google Cloud earlier in the year. According to the initiative’s reported outcomes, the move supported better personalization and helped lift average order value (AOV), while also reducing the time required for product launches.

The next steps focus on scaling the approach: PUMA plans to explore Imagen 3 (Google’s newer text-to-image model) to extend campaign creation, and it intends to expand the use of Vertex AI Search for Retail across additional subsidiaries to strengthen discovery, relevance, and conversion performance.

Source: geekwire.com

Thomas Kurian Google Cloud CEO portrait. This image features a man in a blue button-down shirt, smiling directly at the camera against a plain white background. It is a clean and natural portrait, free from any distracting elements or text overlays. This makes it an excellent choice for an inline blog image slot requiring a professional and clear depiction of Thomas Kurian.

On the analytics side, PUMA is using Google Cloud’s machine learning capabilities through BigQuery to deepen customer engagement via advanced audience segmentation. The approach uses first-party data to build custom ML models, generate predictive insights, and attribute conversions more accurately across touchpoints.

Reported performance gains from this segmentation work include a 4.6% increase in conversion rates, a 6% increase in AOV, and a 149.8% increase in click-through rates for the top three audience segments compared with other advertising targets.

Key Technologies and Partners for the PUMA AI Creator

What Other Brands Can Learn (Without Copying the Jersey)

The jersey story is specific, but the mechanics generalize. If you strip away the football context, what remains is a repeatable pattern: creation + feedback + personalization + infrastructure. Here are the lessons that matter if you’re thinking about AI in commerce.

• Start with a use case people care about: fans don’t generate designs because AI exists; they do it because identity and community exist.
• Design constraints improve outputs: a 300-character prompt limit reduces noise and makes intent clearer.
• Make participation sustainable: credits + voting rewards keep users engaged after the first “wow moment”.
• Engineering is the product: queues, scaling, and GPU efficiency decide whether the experience feels premium or broken.
• Plan governance early: generative systems need moderation, prompt filtering, and brand-safe boundaries (especially at global scale).

Conclusion

The PUMA AI Creator is a clear example of what happens when generative AI is treated as more than a novelty. PUMA used AI to pull fans into the creative process, but the real achievement was operational: shipping a global, high-volume, AI-driven experience that stayed responsive, measurable, and cost-aware.

In ten days, the campaign produced a level of participation (designs + ratings) that most brands only dream about — and it did so while showcasing a modern AI stack: containerized generation on GKE, optimized performance on Nvidia H100 GPUs, and personalization via Vertex AI. If this is the direction retail is moving, the lesson is simple: the future of e-commerce isn’t only “recommendations”. It’s interaction — and increasingly, co-creation.