An AI-powered image generation platform
Salesforce AI Research
My role
UI design
Information architecture
Results
Pixel Playground was showcased at Dreamforce '23. Used as a foundation for the development of AI image analysis for Salesforce's main platform.
Pixel Playground: an exploration into generative AI
While developing xGen, Salesforce's foundational LLM, the AI Research team sought to create an experimental platform to test and improve the model's generative capabilities, accuracy, and guardrails. My team was responsible for the design and development of this multimodal platform, named Pixel Playground.
Pixel Playground uses natural language prompts to perform a variety of image related tasks, including image generation, object removal, and stylization, among others.
With multiple functionalities to choose from, Pixel Playground needed an intuitive UI
After a discussion to understand the program's design requirements and the breadth of the Pixel Playground's capabilities, I began wireframing potential designs for the platform's main pages.
While these initial designs were serviceable for the platform's simpler functions, it became clear that the interface would become cluttered when set to perform more complex tasks, such as advanced image editing.
After receiving similar feedback from the larger research team, I began to explore ways to simplify the entire interface.
The next iteration had the model's capabilities consolidated into three categories, with interfaces specific to each category, such as a two-column format in image editing to allow simultaneous exchanges with the LLM while viewing image changes in real time.
While this setup was an improvement in usability, our team observed confusion from first time users during internal usability testing due to the multiple interfaces.
To answer this issue, I removed the top bar entirely, made the sidebar a history of previous conversations, and housed all editing modes in the input field, which can be called upon with the '/' key. This allowed the interface to remain the same, regardless of the selected function.
With approval from our research director, I proceeded to create flows of each of the model's capabilities, such as the quick edit feature below. I used these to communicate the ideas of each function to the development team.
Finally, I created an instructional guide with visual examples to better describe the platform's abilities.
Mobile
With the desktop version completed, I formatted the interface to be compatible on mobile devices using Pixel Playground in a web browser. This would allow attendees at the upcoming Dreamforce expo to experiment with the platform following the presentation.
Conclusion
The completed prototype was presented at Dreamforce 2023, in which attendees were encouraged to download the mobile version to interact with it themselves.
While the project was only initially created to demonstrate the capabilities of Salesforce's in-house LLM and deep learning library, the advancements the project made in computer vision and image analysis contributed to the development of Agentforce, Salesforce's autonomous agent platform which was introduced the following year.
© 2026