top of page
Meta Generative AI • 2023
Bringing Generative AI to Advertisers with Background Generation
Background Generation is one creative optimization that was introduced to the Ads Manager experience to help businesses create visually captivating performant ads with less effort.
Project Overview
Background Generation simplifies ad creation by giving advertisers the ability to generate background variations for product ads across Instagram and Facebook. This project had 2 work streams; the Ads Manager experience of how advertisers would use this new features, and the Model Improvements of how the outputs of these generations would function. This case study outlines the process and strategy for model improvements; generating better outcomes to increase advertiser adoption.
My contributions
I led the end-to-end design and strategy for the model improvements of background generation; specifically strategized with engineers on how to improve the model to generate better outputs, implemented design standards and principles to focus on quality and scalability, and collaborated with designers, engineers, and leadership across the Ads Manager experiences and other Generative AI initiatives.
What do we already know?
In early 2023, AI Sandbox launched as a "testing playground" for advertisers to try out new generative AI powered ad tools like text generation, image outcropping, and background generation. Background generation received the lowest adoption, due to quality and relevance issues.
1. Quality is Key
Advertisers look for high resolution outputs that have a seamless integration of products into the generated backgrounds.
2. Relatedness is Good
Advertisers want outputs to align and communicate key information about the product and brand.
3. Product is the Star
Advertisers want the backgrounds to be subtle, allowing the product to be the focal point in a way that looks natural and well integrated.
Who are our advertisers?
Small retail and ecommerce advertisers in lean organizations. They often prioritize direct response campaigns over brand and are performance driven.
Advertiser problems
1. Advertisers perceive product ad creatives as generic and boring, but have limited options to customize.
2. Advertisers struggle to create performant ad creatives at scale.
Current product ad on FB Feed
Current product ad on IG Stories
Hypothesis
By improving the quality of background generation, we can have more outputs that meet advertisers needs and align with customers expectations, leading to advertiser adoption and an increase in ad performance.
Advertiser Adoption
In order to improve the model to generate higher quality outputs, we needed to first define what a high quality output looks like. Through extensive research; talking to customers, creative professionals, and investigating performant ad creatives, a team of designers and I across GenAI were able to come up with a framework of how we can define and work towards high quality generations.
Once we had a foundation and principles to abide to for quality, the next step was to dive into the technical structure of how the model worked. This gave a sense of where in the model structure certain improvements could be made and what would be feasible. The 2 opportunity areas were correctness and relevance.
How might we improve the correctness?
Define "good creatives," adding them into the model and improve the feedback loop to train more quickly. The more usage and feedback the model gets, the better it will become.
1. Image Scoring: Set a standard approach to rate quality that can feedback back into the model. Looking at things like color contrast, segmentation, drop shadows, and lighting to determine quality scores. With each rating, quality will improve over time.
Improved the rating system: Quality can be subjective so I incorporated the foundation and principles as an educational piece for reviewers to reference when determining what generations were high quality or not.
2. Quality Questionnaires: There wasn't a ton of usage with Background Generation as it was only used as a "testing playground" with AI Sandbox. With the help of annotators, more and more generations could be assessed. A Quality Questionnaire asked specific questions related to the framework of defining good quality and get feedback at a much higher rate.
Improved the questions and tasks: Incorporated specific questions and tasks to be more quality focused.
Pushed to have more relevant reviewers: Annotators that were completing the Quality Questionnaires were a 3rd Party source, with no background on advertisers or creative knowledge. In order to assess creative quality, we needed relevant people to complete tasks. I worked with the Design Director to help get support from a creative agency to participate in Quality Questionnaires.
How might we improve the relevance?
Improve prompt generation to effectively communicate what should be generated base on categories and themes.
A prompt is a text prompt, used to provide guidance to the model on what types of background should be generated. In order for a prompt to be generated, categories and themes have to be identified. Categories are manually added for each product and themes are identified.. Themes help determine the desired background type and compliment the product and its relevance. Currently, there's a set of predefined background themes and prompt templates that are used to generate prompts for each product.
Problems
1. There's no clear hypothesis used to determine which themes will help performance
2. Prompts are fixed, limiting the variations and ability to produce relevant backgrounds
3. Generations aren't taking into account brand guidelines
Solutions
1. Improve themes by introducing scenes and styles
2. Move to a dynamic prompt to create more diversity
3. Bring in advertisers brand by leveraging historical ads
.png)
Solution 01
Improve themes
We currently focus on Abstract and Lifestyle themes that determine the types of backgrounds that are generated. Abstract themes are simple, focusing on colors and patterns with the product overlaid on top, while Lifestyle themes are a more natural integration of the product within a background. By expanding themes to include other components like environments, surfaces, and styles, we can help train the model to deliver more variants that are relevant, contextual, and on brand.
Being more specific tends to generate higher quality outputs. The diagram below shows how a single product can have a lot of diverse and high quality generations when we incorporate components like environments, surfaces, and styles.
Solution 02
Introduce dynamic prompts
Currently, there's only 3 fixed prompts per product category, where the product category is the only thing that’s dynamic. This means that the environment and styles we reference are pre-defined, posing an issue when a product may not make sense with the suggested style or environment, limiting the variations that can be generated.
With the dynamic approach, the prompt can have placeholders for things like the surface it’s placed on, the environment it’s in, or the styles, the diversity, relevance, and overall quality increases.
Solution 03
Incorporate brand characteristics
Introducing components will create more variations and opportunities for higher quality and relevant outputs, but there could be a risk of these not aligning with advertisers' brands or having similar outputs as their competitors.
Instead we can reference past ads to give insight into what styles can be incorporated. These styles would be specific to that advertiser, creating more personalized outputs that fit the advertisers brand.
Background generation began rolling out to all advertisers in October 2023. The model improvements are continuously being updated to improve the quality and increase advertiser adoption.
bottom of page