At Polarr Next, we have a dedicated team producing high-quality, pre-trained AI Styles that cater to a broad spectrum of user aesthetics. In this blog, we want to share the behind-the-scenes process of creating these Styles, which can be helpful for anyone aiming for perfection in their own AI Styles.
Market Research
Every AI Style in Polarr Next is the outcome of thorough research. We analyze current popular color styles across various domains, including social media, photography exhibitions, and fashion magazines. For each identified style, we then analyze color trends and compile data. For example, when we notice a trend toward warm-toned editing styles, we study the color characteristics of different warm-toned images, including skin tones, foliage, land, flowers, and jewelry.
Outside market data, we engage with photographers on a daily basis. We actively participate in key industry events such as WPPI and regularly engage with photographers to understand their aesthetic needs. Though the demographics of our mobile apps aren't exactly the same as professional photographers, we also leverage data from Polarr's millions of monthly active mobile users to understand their editing preferences by talking to popular Polarr filter creators. By examining the usage frequency and adjustment ranges of different tools in Polarr Next and Polarr mobile apps, we capture the latest color trends and user preferences. This comprehensive research provides robust data support for our AI Style design.
Designing the Look
After selecting popular color styles, we begin the AI Style design process without using AI. Instead, we start with generating a Polarr Next base preset. For each preset we want to produce from our market research, we determine various color editing parameters to ensure each adjustment aligns with the style's color characteristics and user habits. The designed base preset is applied statically to a test set of around 100 images to check initial results. We also test the preset on specialized skin tone color cards to ensure no significant color deviation in skin tones. For skin tones, we use a library of portraits with varying skin colors and conditions. We perform skin tone testing at this stage to make sure there aren't any outlier parameters in the preset that can cause abnormal rendering of skin tones while we train AI Styles in the next phase.
AI Training
Building on the manually designed base preset, we train our AI using the same process of marking reference photos as you do in the app. We have a training set of 1,000 images, including photos from approximately 30 weddings. This diverse set covers over 100 different scenes, allowing our trained AI Style to perform well across different environments. We believe wedding photography is ideal for training because of the wide coverage of people, skin tones, lighting conditions throughout the day, and variations among indoor, outdoor, candid, and still objects. In fact, wedding photographers who shoot full-day weddings face the most challenges from varying light conditions and color temperatures.
We train the Style by applying the base preset we designed previously to each of the 1,000 images, then meticulously adjust lighting and colors to maintain consistency among all scenes, and mark reference edits for all the images. The entire process usually takes 1-2 days for one colorist and produces tens of thousands of slider adjustments for the AI to understand how the Style should work to maintain consistency across scenes.
Performance Testing
To ensure our AI Style performs under all common as well as some extreme conditions, we create a separate test set of 1,000 images from another 30 weddings not in our training set. This separate testing set is completely independent from the training data set and includes more challenging photos that are underexposed, overexposed, mixed color temperatures, extreme color temperatures, monochromatic, and backlit conditions. We then apply the AI Style against the testing set to validate its performance in diverse conditions.
Iterative Optimization and Evaluation
The result of our testing typically does not look ideal on the first try. We review the testing results in a small committee of colorists, examining overall color consistency, aesthetics, and tones across the entire 1,000 testing images, as well as particular underperforming images and what the AI did to those edits. We then discuss revisions needed to improve the AI Style. Sometimes these revisions could be quite drastic and result in a full redo of the AI Style. More often, we will go back to the training set, adjusting the training photos with different parameters to ease the inconsistencies observed in the testing images. For example, we might use the HSL tool to add an orange tone instead of using curves.
The design, AI training, and testing process can be repeated multiple times until the final result achieves unanimous approval from the evaluation groups. After approval, we then produce sample images and make the AI Style available on Polarr Next's explore page. Once the AI Style is shipped, we can analyze how often a Style is opened, used, and exported. This behavioral data allows our team to validate our hypotheses on user trends and make revisions to the AI Styles if necessary.
Your Move
Now that we've shared our creative process for creating pre-trained AI Styles, you might wonder how it differs from the AI Style training you perform in the app yourself. The answer is there is no difference except for the quantity and diversity of reference edits we're making upfront. Because we need to design our AI Style to work on all types of photo shoots, we perform thousands of reference edits to cover all those scenes and scenarios to train the AI Style. However, when you make your own AI Style from scratch, you typically do not need this many reference edits to start making your photos look consistent and can train the AI incrementally as you fix its mistakes in real-time.
The general principle is the same: the more diverse reference edits the AI sees for a particular aesthetic, the more accurate it will become at making diverse scenes consistent. You could in fact start editing with our pre-trained AI Style, then continue to add more reference edits, making it even more perfect for scenes and scenarios we didn't cover in our training phase.
Closing Thoughts
We take pride in our team's expertise and experience in creating AI workflows to realize the aesthetic visions of our users. Every pre-trained AI Style in Polarr Next is the culmination of our team's wisdom and effort. We are committed to continuous innovation and refinement to provide users with the best photo workflow experience. We hope you discover more surprises and inspiration with Polarr Next.
