State-of-the-“art”: machine learning for graphics and UI designs
Yuexi (Tracy) Chen reflects on her 2022 HCIL Symposium tutorial on ML and design
Before starting my summer internship at Adobe Research last year, I read several papers about machine learning for design. While neither of my intern projects focused on this, I found the research fascinating, so at the HCIL symposium 2022, I gave a 1-hour talk on the research progress of ML for graphics and UI design. Inspired by one of the attendees — Prof. Hal Daumé III’s Tweetstorm — I wrote this post to share a skimmable version of the talk.
Table of contents:
Part 1: Find design examples/inspirations
- by Cascading Style Sheets (CSS) properties
- by descriptive words
- by visual properties
- by images
- via a chatbot
Talk format:
Via an informal literature review of selected papers published at top-tier machine learning/human-computer interaction conferences/journals, I focus on two topics: “find design examples/inspirations” and “generate layout automatically.”
The symposium talk was hybrid: ~40 attendees joined the talk, and half joined virtually; some were designers or students in design-related majors. Thanks to Slido, I collected primary data about audiences’ preferences on features enabled by machine learning/data-driven methods.
Part 1: Find design examples/inspirations
Motivation: Designers routinely seek inspiration. Findings from a study with 24 UI/UX designers found that most searched keywords on Google, Behance, Pinterest, etc., and collected search results in different aspects: layout variations (N=22), font styles (N=14), color palettes (N=11).
Pain points:
- It’s difficult to find the right textual description to search for
- It’s time-consuming to find good examples
- Search results do not meet the design requirement
How can machine learning/data-driven methods help?
1.1 search by Cascading Style Sheets (CSS) properties
Cascading Style Sheets (CSS) is a style sheet language decorating web content. Users can choose any part on a webpage and inspect information about font family/size, color, padding, etc. (Read about how to do it.)
Researchers downloaded 100k+ web pages and built a giant database. Users can search design examples in the database by CSS properties, e.g., font size, widget position, and global page layout. Here are some search results:
What attendees said:
One of the attendees expressed concerns about searching by CSS properties, as the same widget could be named differently across webpages (e.g., carousel vs. slider).
1.2 Search by descriptive words
Researchers have also mapped descriptive words with Document Object Model (DOM) and CSS properties. Users can type descriptive words directly; after translation, the search engine will retrieve results based on the definition of CSS properties.
What attendees said:
Attendees shared common keywords they use; examples from designers in attendance included:
minimal/minimalism/minimalistic
neat/clean/simple/easy
colorful/bright
cheerful/fun
dark
1.3 Retrieve images by visual properties
In the pre-deep-learning era, retrieving similar images by visual properties achieved reasonably good results. For example, by combining color histogram and Histogram of Oriented Gradients (HOG), researchers have found stylistically similar infographics.
1.4 Retrieve images by image embeddings
Neural network-based methods allow users to search for similar UIs by high-fidelity screenshots, wireframes, sketches, and even partial sketches.
Although the above methods take different inputs, they adopt the same approach: learning image embeddings (a low-dimensional vector representation) and performing nearest neighbor searches.
1.5 Retrieve images by chatbots
Chatbots could help users find design examples in a database via conversations.
Attendees’ preferred methods:
23 attendees voted for their preferred methods to find design examples/inspirations (note: multiple selections were allowed; see figure below). Searching by wireframes was the most popular (65%), followed by high-fidelity screenshots (57%), while searching by chatbot (26%) and stylistic keywords (22%) were less popular. This is likely because designers found keyword-based search results to be too broad.
Part 2: Generate layout automatically
Motivation: When seeking inspiration, most designers (22 out of 24) collect layout-related examples. If machine learning could automate the layout creation process, that would be a great time saver. Below I discuss options for doing this.
2.1 Generate layout based on constraints
Scout (demo video) generates feasible layouts given constraints specified by users. Constraints could be non-overlapping, minimal sizes, alignment, visual hierarchy (paddings within groups), emphasis (can’t decrease the size of an important element), etc.
What attendees said:
Designers shared typical constraints they encountered, including:
information hierarchy
branding requirements
headers and footers
2.2 Generate layout based on “energies” (design guidelines)
DesignScape (demo video) generates layouts by minimizing design guideline violations (“energies”). Design guideline violations include alignment violations, balance violations, overlap, etc.
2.3 Generate layout based on keywords/categories
Researchers have trained Generative Adversarial Networks (GAN) to generate magazine layouts based on images, text, and magazine categories. Below are different magazine layouts generated for the same image but with different keywords.
2.4 Generate layout based on spatial relations
Researchers have also modeled spatial relations among objects like graphs and trained Graph Convolutional Networks (GCN) to generate designs satisfying such relations.
2.5 Generate layout with background and embellishments
Most ML methods generate layouts using only given assets. A prototype called Vinci could suggest background and embellishments (demo video). Vinci views the design process as linear design sequences and leverages a sequence-to-sequence Variational Autoencoder (VAE) to select background and embellishments.
Attendees choice:
21 attendees voted for their favorite layout generation methods. Vinci garnered the most votes (62%), followed by spatial relations (48%). Full results are below.
Takeaways:
- Some important ML techniques don’t lead to popular features, e.g., search design examples by chatbots.
- The same ML component could lead to features with varied popularity, e.g., searching by wireframe is more popular than searching by high-fidelity/sketches.
- Designers are eager to adopt such techniques! One designer even asked whether it’s possible to download one of the “tools.” Unfortunately, it’s only research code: no server, no front-end 🤷♀
Further reading:
Some papers don’t fall into the above two topics but are still interesting to read:
