How AI‑Generated Images Supercharged Baidu App’s Ad Growth

Impossible Task?

Network advertising requires thousands of personalized creatives each week, a workload that manual design cannot sustain. Traditional methods struggle to increase both quantity and visual quality while keeping costs low.

MEUX’s operation designers partnered with product and R&D teams to use Stable Diffusion, generating massive image assets that surpass manual production in both volume and quality.

From Rookie to Pro

Early attempts with Stable Diffusion often failed, especially when generating images for diverse search queries such as "home meatballs," "how to refuel a car," or "spicy crayfish recipe." The team refined prompt engineering, selected appropriate models and LoRA weights, and iteratively tuned parameters to achieve satisfactory results.

For particularly difficult queries, they employed img2img, ControlNet, and other plugins, noting that many Chinese‑specific concepts still challenge existing models.

Hand‑in‑Hand Mass Production

Initial small‑scale production involved selecting a handful of queries that AI handled well, crafting detailed prompts, and generating thousands of images for online deployment. However, Baidu’s search queries number in the hundreds of thousands, making manual prompt creation infeasible.

The solution was to classify queries by verticals and create prompt templates that control color, perspective, texture, and emotion without specifying the main subject. Designers supplied subject‑specific keywords programmatically, allowing Stable Diffusion to batch‑generate images.

Examples include using realistic models like Realistic_Vision_V5.0.safetensors for manufacturing scenes and Asian‑beauty‑focused models for medical‑beauty content, with positive and negative terms to avoid distortions.

After establishing base‑image templates, the team designed generic layout templates for text and call‑to‑action elements, then automatically overlaid them onto the AI‑generated backgrounds. Different verticals received tailored layouts—for landmark images a top‑bottom structure, for wide‑angle nature shots a centered layout with bold copy.

Road Ahead

During large‑scale deployment the team observed that more material does not always equal better performance; visual appeal and copy quality are equally critical. They refined templates for categories like tourism, adjusting color tones and perspectives to boost click‑through rates.

Copy generation remains a bottleneck, especially for abstract queries where strong headlines drive engagement. Designers continue to experiment with AB testing to identify “good material” characteristics and plan to explore AI‑assisted copy generation.

Overall, AI tools have become a foundational skill for designers, requiring strong business sense to translate technical capabilities into commercial value. Ongoing research aims to deepen human‑AI collaboration and turn more impossible tasks into reality.