Shu Zhang
中文
← Selected work
02 — AI & Creative Systems

Betting on Layout, Not Just Images

Designing the rule system that let Microsoft Advertising partners generate editable, production-quality banners across top display ad sizes from a single asset collection — in late 2023, before AI design workflows went mainstream.

Role
Design lead — AI banner generation logic & rule system
Partners
ML Scientists, Platform Engineering, Product, Partner Advertisers
Timeline
Q4 2023 – Q3 2024 (backend Q2, web experience Q3)
Outcome
Shipped in Microsoft Advertising creator tool

Partners needed banner ads across all top display sizes. Each resize meant manual work: swapping copy, repositioning elements, adjusting layouts. What used to take a designer a full day could now happen in minutes — but only if we could teach a machine the design rules, not just generate pictures.

The problem: production reality vs. AI hype

By late 2023, most AI creative tools — Canva, Stable Diffusion, Midjourney, Microsoft Image Creator — focused on generating imagery. Beautiful outputs, but static. Uneditable. That didn't solve the real workflow problem: advertisers and publishers still needed to produce dozens of banner sizes, each following brand guidelines, each requiring manual adjustment.

We used to ask designers to create these manually, or hire agencies. Communication took time, quality varied, and updating all sizes was a day's work.

Partner feedback, early discovery

The pain wasn't generating one pretty image. It was adapting one ad concept across all top display sizes — portrait, landscape, square — while respecting partner brand guidelines and producing editable output they could still customize.

The bet: teach layout logic, not templates

My thesis was different from the industry direction. Instead of generating static images, we'd build a system that understood design rules and generated editable layouts. Not "here's a picture" — here's a structured design you can still modify.

The shift: from templates to logic. Define how a banner should adapt, not what it should look like.

What I designed: the rule system

I owned the banner generation logic — the decision tree the AI would follow to create production-quality output. Working closely with ML scientists and platform engineers, I defined the design system as executable rules:

The key insight: instead of designing templates, I designed the logic that generates templates. The system needed to understand layout, not just reproduce it.

Four-step banner generation rule system showing: 1) Segment selection by ratio - vertical, square, and horizontal segments grouped by aspect ratio ranges for desktop and mobile; 2) Asset combination by segment - different asset combinations like logo+CTA, logo+title, title+CTA, and logo+title+CTA for each segment type; 3) Constraints - sizing rules for overall dimensions and headlines, position rules for CTA placement, and accessibility requirements for logo size and text contrast ratios; 4) Ranking - alignment, sizing, and logo quality scores used to rank generated layouts
Design rules as executable logic — instead of templates, I defined how layouts should adapt across aspect ratios, asset combinations, and brand requirements.
Multiple Pure Leaf tea banner ad variations generated from the same asset set, showing intelligent layout adaptation across different aspect ratios and sizes - portrait, landscape, square, and skyscraper formats. Each banner maintains brand consistency with the Pure Leaf logo, product imagery of colorful tea bottles, headline 'REAL VARIETY, REAL BREWED', and red 'Shop Now' CTA button, but with layout logic tailored to each format's dimensions and proportions.
Rule system in action — the same asset collection intelligently adapted across multiple aspect ratios and dimensions.

Cross-functional collaboration

This wasn't a solo design effort. I worked in a tight loop with:

I also scoped the end-to-end creator experience — from web UI to generation logic — but the scope was too large for one designer. Another designer joined to own the product UI/UX while I focused on the generation system.

Three concept screens for the Creative Creator tool showing different interaction models: Gorilla mode with bulk asset grid upload, Chameleon mode with style simulation and layered composition preview, and Crow mode with modular assembly showing headline text field and configurable layout controls. Each concept explores different approaches to asset input and banner generation.
Early vision for the end-to-end creator experience — from asset input to multi-size generation. This strategic framing guided the team, while another designer later led the product UI/UX implementation.

What shipped

The creator tool launched in the Microsoft Advertising platform. Advertisers, publishers, and brand marketing teams could now generate production-quality, editable banners without design expertise or agency delays.

The impact: What used to take a full day of manual work — updating dozens of banner sizes one by one — now happened in minutes, with editable output partners could still customize to their needs.

Microsoft Advertising creator tool interface showing the banner generation workflow. Left sidebar displays asset management with uploaded logo, recommended logo variations, source image, and recommended images. Right side shows generated banner previews in multiple sizes - 970x250px billboard, 728x90px leaderboard, 320x50px mobile banner, and 300x250px medium rectangle - all using the same Adventure Works brand assets with 'Conquer Rougher Terrain' messaging, demonstrating intelligent layout adaptation across formats.
The shipped Microsoft Advertising creator tool — where partners generate production-quality banners using the rule system.

Why this mattered in 2023

While the industry was racing to generate prettier images, we were solving a different problem: removing manual toil from production workflows. The differentiator wasn't better pictures — it was editable, rule-based layouts that understood design logic.

That's a bet on AI as a tool that learns design principles, not replaces designers. The same thesis shows up in my later work on AI-assisted content operations: automation should remove repetitive work, not creative judgment.

Honest scope

I owned the generation logic and rule system, not the full product. Another designer led the creator tool UI/UX. My contribution was defining how the system thinks about layout, not every surface a user sees.