Within the last blog, we discussed how AI is used to enhance the invention experience in Monotype. If you have got not read it, please read it here Artificial Intelligence journey at Monotype

We looked briefly at Font Similarity, WhatTheFont, and Font classification. Today, we’ll talk concerning the newest edition of our AI arsenal — Font Pairing

Have you ever ever cooked a meal at home and been delighted to search out the right ingredient to enhance your recipe? Possibly you forgot you had cilantro, crushed peanuts, or lemon juice, and it’s just the thing you needed to raise your dish. Or perhaps, you were getting dressed for the day. You chose your favorite pair of denim but weren’t sure which shirt would match. You considered your mood after which the destination. You finally selected a warm, seasonal sweater. In either scenario, you created a fresh and distinctive flavor, an enhanced experience. And consequently, a sensation of ease and satisfaction washed over you.

That is what finding the right font pairing looks like. These are the every day decisions you make about harmony and contrast. These pairings are thoughtful, fastidiously curated designs. Like pairing anything, it’s a subjective practice with infinite possibilities. There are various good decisions, some which might be bad, and a couple of which might be exceptional.

Now imagine if you happen to had an AI tool that considered your expertise and emotions as your personal assistant. It could offer exciting, delightful ideas, and increase the speed and quality of your creative decisions. These suggestions could even act as a soundboard or a source of inspiration. We’re using AI to extend the depth and breadth of our font pairing suggestions.

Even the experts within the typography world don’t at all times agree on the perfect approach in selecting a pairing font. Making a general solution that caters to typographic experts in addition to a beginner could be a tricky task, and we’ve tried to tackle this challenge a couple of times ourselves.

Up to now, we’ve attempted to create some analytical rules to guide our pairing decisions. Nevertheless, we found that these rules didn’t at all times work across all fonts, even in the event that they were grouped together under the identical category. This was on account of subtle differences within the skeleton and structure of every font.

While there are many font pairing tools available now, most of them are hand-curated and can be found just for bestselling fonts. And while some ML-driven font pairing tools exist, they could not at all times satisfy the attention. So, we decided to explore some recent techniques to search out higher solutions for pairing fonts.

Ideation

Before solving, we took a step back to research the issue and develop a transparent process. We broke down the delivery/research into manageable steps, with the goal of solving probably the most essential problem for our audience.

Our journey began with making a model that might understand style, skeleton, contrast, similarity, and other vital concepts that might be the backbone of a powerful font pairing engine.

These manageable phases include:

1. Harmonious pairing for normal fonts
2. Harmonious pairing for daring fonts
3. Script font pairing
4. Contrasting font pairing

Harmonious font pairing in easy terms means selecting two or more fonts that complement one another and work well together. Some things we kept in our mind while deciding the criterion to decide on harmonious font pairings:

Pairing fonts which have different styles can create a harmonious contrast. For instance, pairing a sans-serif font with a serif font can create an interesting and balanced look.
Fonts that belong to different classes and are more distinct of their skeleton and structure add variation and visual interest to pairs.

Examples of harmonious font pairings could also include pairing a sans-serif font with a classic serif font or using a contemporary serif font with a straightforward sans-serif font.

Execution

Once a transparent approach was chosen and agreed upon, we created a scalable solution with the goal of developing a system that might analyze a considerable amount of data and make accurate recommendations in real-time. After careful consideration, we selected to go via the embeddings route to know font structure.

are a kind of machine-learning algorithm that may analyze the structure and discover patterns. It’s a strategy to represent data, akin to text or images, in a numerical format that could be easily processed by machine learning models. Essentially, embeddings convert complex information right into a format that a pc can understand and work with. In Monotype, we have already got a few AI engines that work using embeddings.

Ultimately, we took the assistance of our studio team to create a sample dataset for training the model. Then to create the answer, in-house type experts analyzed the information expanded it, and cleaned the information before training a state-of-the-art machine learning model on it. While the initial accuracies were promising, it took a while to fine-tune the model to work effectively with the fonts available on MyFonts.

The training process involved feeding the machine-learning model with fastidiously chosen good and bad pairs of fonts and their embeddings. The model then learned to check the embeddings and generate a similarity rating for the fonts. The upper the similarity rating, the higher the pairing of the fonts. The model was also fed extra metadata like font styles and categories, akin to serif and sans-serif fonts, and used this information to create more accurate pairings.

Evaluation the outcomes

After the training process, we tested the font pairing engine on a big dataset of fonts from MyFonts. We evaluated the engine’s performance by comparing its results with the judgments of skilled designers. The outcomes were promising, and the font pairing engine was capable of produce accurate and visually pleasing font combos that were on par with the designers’ recommendations.

We are actually working on advancing the model’s capabilities to incorporate display typefaces because it has the power to know the skeleton and structure of fonts.

Users have overwhelmingly responded positively to the font pairing model launched on the playground. Using playground, the users can experience random AI-generated font pairs, they will like or dislike pairs. We’re generating synthetic data and having type experts fine-tune it while removing errors.

Broadening the model’s capabilities will enable it to offer users with more accurate and diverse font pairing recommendations. We expect that this update will enhance the user experience as pairings might be available on an even bigger inventory of fonts.

This blog has been written by Prince Dhiman(Manager, AI/ML @ Monotype)