Dynamic Descriptive Model: A Scalable Paradigm for High‑Quality Native Creative Generation

Creating advertising creative images is one of the most cognitively demanding tasks for advertisers. The creative image is the medium that conveys product or brand information to consumers and is a prerequisite for successful ad delivery. Traditionally, advertisers spend extensive time using tools such as Photoshop to manually design these images.

With the rapid expansion of internet media, the number of ad slots and their format specifications have multiplied. To serve all slots, advertisers must produce a separate creative for each requirement, and the need for frequent updates (e.g., for holidays or new products) makes manual production infeasible. Consequently, ad platforms provide programmatic, low‑cost, large‑scale creative generation methods, commonly known as programmatic splicing creatives . Designers create static universal templates with placeholder areas for product images and copy; an algorithm then fills these placeholders with collected assets.

Although splicing creatives are cheap and automated, they often suffer from a generic, stitched‑together appearance that lacks the native feel of manually crafted designs. To bridge this gap, we propose a new paradigm called the Dynamic Descriptive Model (DDM) , which leverages visual perception technology and an intelligent design language to produce high‑quality, native‑looking creatives at scale.

System Overview

The DDM pipeline consists of four main stages:

Material acquisition: automatically crawl product pages for images, copy, and other assets, then filter out unsuitable materials using a classification model.

Visual perception: apply deep‑learning methods to build a large‑scale material perception library that captures multi‑dimensional attributes of each asset.

Intelligent design language: abstract designers’ experience into a descriptive template language, extending standard SVG with abstract objects and operations.

Generation: match descriptive templates with suitable material sets, convert abstract operations into concrete parameters, and render the final SVG, optionally via a real‑time rendering engine.

Material Mining & Filtering

We obtain high‑quality assets at low cost by using visual perception models to extract images from product detail pages (e.g., main images, detail images, video frames) and filter out low‑quality or unsuitable items.

Material Perception

Understanding the material is essential. Our perception system analyzes assets along three dimensions:

Layout attributes : element detection (objects, models, logos, text) and image extension using an Outpainting GAN to adapt to diverse slot sizes.

Subject segmentation & shadow generation : fine‑grained matting (down to hair strands) and automatic shadow synthesis based on lighting direction.

Style / color matching : scene classification and detailed region‑wise color analysis to ensure harmonious visual composition.

Intelligent Design Language

We extend SVG (Scalable Vector Graphics) with abstract objects (semantic image descriptors) and abstract operations (composable actions such as selection, placement, mirroring, extension, and color adjustment). These abstractions allow a template to describe *what* should be done rather than *how* to draw each pixel.

Matching & Production

During generation, the system parses a descriptive template, filters materials based on abstract object conditions, converts operation parameters to concrete values using the perception results, and modifies the SVG accordingly. A set of rendering commands can be invoked via URL to obtain the final creative in real time. This approach enables massive, flexible production while preserving visual fidelity.

Business Impact

Deploying DDM has yielded noticeable visual improvements over traditional splicing creatives and click‑through rate gains ranging from 5 % to 80 % across various marketing scenarios. The Alibaba Mama creative platform now hosts over 2,000 descriptive templates, supports dozens of ad formats (banner, splash, feed, etc.), and generates roughly 80 million creatives daily across multiple media channels.

Template Editor & Demo

An interactive template editor has been built jointly by front‑end engineers and designers, allowing users to compose abstract components, preview batch outputs, and iteratively refine designs. A live demo is available at https://chuangyi.taobao.com/pages/ddm .

About the Team

The Alibaba Mama Creative Algorithm team focuses on AI‑driven upgrades for advertising creation, covering computer‑vision, video, copy generation, and more. Their research has been published at ICCV, AAAI, ACM MM, WWW, EMNLP, CIKM, ICASSP, among others. They welcome talent in CV, NLP, and recommendation systems.