In a recent interview with SafetyDetectives, Bartol Freskura, co-founder of TensorPix, shared the startup’s journey since its inception in 2020. TensorPix, focusing on image and video enhancement, overcame initial challenges in media restoration by pioneering cloud-based AI processing, ensuring scalability without the need for powerful local hardware. The interview highlights TensorPix’s standout features, including a user-friendly experience, a comprehensive enhancement solution, and a forward-looking growth roadmap. Freskura discusses how AI addresses upscaling challenges while maintaining quality and outlines common issues faced by content creators. He emphasizes the importance of data privacy and foresees significant advancements in the field, predicting improved output quality and potential price reductions in the next 2-5 years.
Can you talk about your journey and what led you to start up TensorPix?
TensorPix began in 2020, just before COVID hit. It was founded by me and my co-founder Branimir. It all started when he approached me with an idea about the abundance of low-quality media archives in broadcasting. For instance, Croatian Radio and Television has a full archive of files that can’t be utilized due to their quality.
The cost of restoration would be prohibitively expensive because it involves manual labor by a very skilled workforce. It’s also time-consuming, so traditionally, restoration was only feasible for top-tier movies and series.
He reached out to me because of my background in computer vision and machine learning. At that time, I had around five years of experience in machine learning, particularly in developing B2C applications with tens of thousands of users.
In my previous roles, it was crucial to focus on the performance of machine learning models, not solely on their accuracy or quality, but also their efficiency, as they needed to be fast to serve 10,000 users simultaneously.
In 2020, Branimir asked if I could apply AI to automate the video restoration process. I delved into academic research, built a prototype, and it turned out that the technology was somewhat promising, though not perfect.
We presented our prototype to Croatian Radio and Television, but that’s where things stalled because, being a large national company, they required a tender to do business with them.
After that, we shifted our focus to private companies, which also didn’t work out because, at the time, no one was interested in our technology. Our last resort was to create a self-service B2C application where anyone could create an account, upload their video, click a few buttons, and have their video enhanced within minutes.
We invested considerable effort into this, and after a year and a half, we began to gain some traction. The user base was growing by 20% month over month. The last time I checked, we had over 700,000 registered users on the platform.
So I would say we’re doing quite well. We’ll continue to make improvements.
What are the key features that make TensorPix stand out in the field of image and video enhancement?
TensorPix distinguishes itself in the image and video enhancement industry through several key features:
- Cloud-Based AI Processing: One of TensorPix’s standout features is its use of cloud computing for AI processing. Video processing, especially when combined with AI, is one of the most computationally intensive tasks. To run AI algorithms effectively, one typically needs high-end hardware, particularly GPUs, which are crucial for AI processes. TensorPix addresses this by processing all AI tasks in the cloud, which has virtually unlimited capacity, even for GPU resources. This scalability means that when a user needs to process a video, additional GPU power can be allocated seamlessly, eliminating the need for the user to have powerful computing hardware at home.
- User Experience (UX): TensorPix prioritizes a user-friendly experience, focusing on simplicity to cater to its average user, who is more likely to be an amateur than a professional video editor. The platform is designed to be intuitive, with a streamlined process from sign-up to video enhancement that takes only a few minutes.
- Comprehensive Enhancement Solution: Unlike other products that may specialize in only upscaling, denoising, or framerate improvement, TensorPix aims to be a comprehensive solution for all video quality issues. The goal is to eliminate the need for multiple programs by providing a one-stop-shop for video enhancement.
- Growth and Roadmap: While TensorPix does not yet solve all video quality problems, it has a clear roadmap for future development and improvement, indicating a commitment to ongoing enhancement and innovation within the platform.
How does AI-based video enhancement technology address challenges like upscaling and improving clarity without sacrificing quality?
Upscaling is not a new technology; many people think it’s something recent, but it has been done since the first videos were created. You can open a program like Photoshop or Premiere Pro and increase the resolution of your video.
The significant difference lies in the algorithm used. The default algorithms today are called bilinear or bicubic, which are very fast and can run on any hardware you want, but the quality isn’t so good. You’ve probably noticed if you take a low-resolution video and stretch it to 4K, it will become all blurry and lose details; it won’t look good. This is the core difference between non-AI upscaling algorithms and AI-powered ones.
At TensorPix, we exclusively use AI-powered algorithms because we aim not only to increase the number of pixels but also to maintain all the details, even if you’re upscaling from SD to 4K. So, we address two scales as efficiently as possible.
That’s the biggest difference between non-AI and AI software.
What are the most common issues that content creators face when trying to enhance the quality of their videos?
The biggest problem we’ve seen from our users is the diversity of the source video quality. Professional video editors understand that you can’t get a perfect 4K video from a low-quality source. The problem lies with amateur content creators who don’t understand the technology behind the scenes, and it can be a bit difficult to explain what is possible and what’s not.
Sometimes you have a pretty good source; the quality is okay, but you want to get that extra detail and upscale from HD to Full HD. That’s not too big of an issue. What is a problem is when users want to enhance an older video, taken on a flip phone or with 360 resolution. They come to our platform expecting to get crystal clear 4K video, which is extremely hard.
You must have a good starting point in order to expect some kind of improvement. So, managing expectations is a huge issue that we deal with.
The other problem is that video content creators are often limited by hardware. As I said before, video processing is very compute intensive; you need good hardware, especially if you’re going to use AI.
We’re trying to solve this problem with cloud computing, which gives us infinite resources, and creators don’t need to worry about the hardware they have at home.
What are the considerations for data privacy when using AI to enhance personal photos and videos?
Data privacy is a really hot topic these days, and I think it’s really important, especially for online providers like TensorPix, because we have to take user data, process it, and return it back. That’s one of the crucial aspects of our platform. I think there are like two perspectives here.
From the company perspective they should do as much as possible to protect sensitive data. This includes using advanced and updated encryption algorithms, regulating which staff members have access to user data, implementing multi-factor authorizations, using long and random passwords, and all other good security practices. That’s what we’re doing on our end to prevent any data leaks or hacks.
From the user perspective, it’s really important to always check the privacy policies, terms of agreements, and FAQs. There are companies that explicitly state that they will sell user data to third-party companies. Our terms and agreements explicitly state we don’t sell user data to anyone.
How has the field of image and video enhancement evolved over the past few years, and what trends do you foresee in the future?
The field of AI image and video enhancement is relatively new, so there’s been a lot of rapid progress. Our algorithm from three years ago was much worse than what we have today. We expect significant improvements, and the output quality will drastically improve in the next 2 – 5 years. Maybe you’ll even have the ability to fix old videos from flip phones or very low-resolution videos.
I also believe that the pricing will go down. Right now, video processing is quite expensive because the AI models are heavy to run. However, over the coming years, there should be advancements that drastically reduce the required resources. That will mean we could also reduce the price of our product because our cloud costs will be lower, which would mean lower prices for the end user.