How we built a photo editing AI for the property marketing industry

How we built a photo editing AI for the property marketing industry

How we built a photo editing AI for the property marketing industry

The UK is well known for its grey skies and gloomy days. When a potential buyer is looking for their dream property online, it’s essential to show what a property can look like on the best days of summer. An image must stand out on property portals, such as Rightmove and Zoopla, of which 90%+ of their photos are edited through the human photo editing process that relies on skilled photo editors working lengthy hours to edit imagery. When an image is manually edited in 3 hours compared to 24 hours, it can cost double. Consequently, the human photo editing process is not scalable.

AI (Artificial Intelligence) is becoming more competent in enhancing workflows that typically rely on large human workforces. It can also disrupt the photo editing industry. When we started building Autoenhance.ai in July 2020, we had a dream of creating a powerful AI to instantly enhance photos to the highest quality for the property marketing industry.

In this article, we’ll share some of the challenges we have faced towards building a photo editing AI for the property marketing industry. We hope you will find our achievements inspiring, illuminating and exciting.

Recognising AI’s potential to edit property photos

The potential of AI to edit property photos quickly and to a high quality will enable estate agents to list their properties faster and potential buyers to imagine themselves in their new homes. After sourcing data of property photos online and paying for them to be enhanced by outsourced photo editors, we trained our first AI-based on Nvidia’s open-source research, on picture to picture generative adversarial networks. Our first breakthrough occurred when we understood that AI has the potential to “enhance” a photograph, like a human.

In these examples, the AI makes colours more vibrant and replaces the sky.In these examples, the AI makes colours more vibrant and replaces the sky.

However, as this method requires a lot of computation, achieving a high-quality image becomes challenging, and the highest resolution image we could output was 1536x1024 (smaller resolution than an HD tv). Our users were requesting higher resolutions, and we could not compete with the human outsource equivalent. We also discovered that our sky replacements felt unrealistic and pixelated. We then found research on a high-quality photo enhancement AI we could adapt for Autoenhance.ai. We trained the AI on high-resolution data (3072x2048 — double the resolution of an HD tv), and it worked with very little compute, making it extremely scalable.

Colours are just as vibrant with our new model as it was with the first.Colours are just as vibrant with our new model as it was with the first.

Our AI trained from thousands of examples, and we can now produce high-quality enhancements which are as good, if not better, than human-edited equivalents.

Overcoming technical issues: Replacing grey clouds with blue skies

We expected that our image enhancement AI could handle sky replacements, as it replaced grey clouds with blue sky. A believable sky replacement is complicated and, even with a competent image enhancement AI, we couldn’t generate believable skies to a high standard.

To address this, we used a separate AI that identified the sky in the photo and replaced it with a high-quality sky image. This should have worked correctly in theory, but we needed the AI to work on high-resolution images (at least 1536x1024, ideally 3072x2048). Although we collected lots of data, we still experienced failures around finer details in images.

As you can see above, the branches aren’t being cut around, which is unrealistic.As you can see above, the branches aren’t being cut around, which is unrealistic.

Our first attempt at sky replacement AI worked for 60% of users images, but we felt this was not commercially acceptable. After further research, we discovered a solution that could reduce our computation and increase our quality of sky replacement. It produced impressive results on complex images with fine details at a high resolution (3072x2048 — double the resolution of an HD tv) and gained user approval.

Our new AI was able to identify fine details such as the branches and create a believable, high-quality sky replacement.Our new AI was able to identify fine details such as the branches and create a believable, high-quality sky replacement.

One of our greatest achievements is the flywheel that we created for our service in areas that needed improvement. Users identify flaws in photo enhancements, then we are notified and collect data to retrain our AI. In the next few months, we will have more images going through our system every day, which improves our AI’s intelligence.

Developing an interface to launch Autoenhance.ai

We needed an intuitive website and web app to present our concept to users, who could upload their images and test our AI. Initially, we offered a monthly subscription and hid our service behind a paywall, yet we found this detrimental to our growth. While potential users were enthused to try our AI, they were turned away by the paywall before trialling our service, as there was little external validation or proof of concept.

As we were still trying to understand our product-market fit, we needed potential users to try our service. As soon as we launched a free trial offer, we experienced a massive increase in users joining our service and uploading images.

We have been onboarding small batches of customers over the last six months. This has boosted our confidence and allowed us to identify critical areas for improvement to our service. Additionally, we were featured in photography forums, such as WGAN (https://forum.we-get-around.com/), which led to a spike in users.

This graph shows the daily number of images and orders we have received in the last six months.This graph shows the daily number of images and orders we have received in the last six months.

What does 2021 look like for Autoenhance.ai?

Creating a photo editing software for the property marketing industry was never going to be straight forward. We needed to understand the limitations of our technology before we could improve it, as well as comprehend who would use our service and how they would interact with it.

We want to obtain real-world usage of Autoenhance.ai to improve our service. In 2021, we aim to grow the number of users joining our platform and allow them to upload photos to our database daily. We have reached out to large property photography businesses to discuss pilots, and we hope to share more details as they progress this year.

Photo editing styles help brands stand out, and everyone has a preference. From vibrant or natural colours to cloudy skies to clear blue, we want to simplify Autoenhance.ai and enable users to leave their mark on photos. We are currently working out how we can offer a custom style service whilst enhancing images to a high quality within a quick turn-around time.

Our goal is to create a service that replicates the human photo editing process. To supplement our basic enhancement service, we are looking into removing rogue objects from images autonomously, such as clutter or bins.

We have experienced a rewarding and challenging six months. In 2021, we are eager to grow Autoenhance.ai and add plenty of features. We are excited to see our service evolve to provide leading photo editing software for the property marketing industry and we are excited to share our developments as we progress this year.

If you are currently outsourcing your property photos and looking for a better solution, then reach out. We will be happy to organise a call and show you how Autoenhance.ai can work for you.

**Click here to register your interest**

HomeAPIPricingBlogDashboard

© Copyright 2021, Autoenhance.ai. All Rights Reserved.