Demystifying AI, algorithms and automation
Skip this section if you are well-versed in computer science lingo. For the rest of you, we’ve got your back. Join us as we explain what it all means.
AI
Artificial Intelligence is the training of computers to reason and find solutions by simulating human intelligence. Many of the things we interact with on a daily basis use AI, including chatbots, SIRI, and the facial recognition on your phone.
You’ve probably heard of machine learning and deep learning. These are both part of the AI puzzle. Machine learning and deep learning are tools to help train AI. For AI to become “intelligent”, it needs to learn. AI is trained with vast amounts of data, which can include anything from images to text, numbers and sounds.
Algorithms
Simply put, algorithms are a set of instructions. Imagine a recipe - your input is all the ingredients, the recipe is the algorithm with specific instructions and the output is your Thai green curry. Sounds simple enough (we promise, it’s not). At least not when we get to writing advanced algorithms for things like AI and automation.
Automation
This one’s easy, right? Not a trick question, automation does what it says on the tin. It minimizes the work you and I (mere mortals) have to do by automating things. Most of the time, an algorithm is needed to set up an automation, but not always. Super handy for speeding things up, there are plenty of tools today that even automate the automation (trippy).
Little did you know, you’re probably already using automation. In Photoshop, automations include things like batch processing. All that is required from little old you is to save your settings and it’s ready to go whenever you need it. Applications like Zapier automate workflows by connecting different tools without having to write any algorithms.
For instance, you may get an email with new images that are automatically saved to Google Drive for you - cheers Zapier.
In layman's terms
From the bottom up - automation runs out a set of pre-programmed rules whenever you press go. Algorithms are the pre-programmed rules themselves. AI learns what the rules are, by intelligently making decisions based on information it’s been trained with. Nifty.
Take for instance when we trained our AI to remove moles on skin. We fed the AI a load of pictures - 65,000 manually sorted snaps to be precise - so it could learn what moles are (not the ones with fur and poor eyesight). Once it knew, we created an algorithm for mole removal. Now whenever we fancy, we can automate this algorithm to easily remove moles willy-nilly, saving time and boosting efficiency. Smart huh?
What does AI photo editing mean for post-production?
We still need humans in post-production. Humans can do a lot of things that AI can’t. When it comes to eating spaghetti, cringing at Dad Jokes, or going trampolining, AI’s got work to do. In the world of photo editing though, AI can be pretty useful. Never fear, using AI doesn't mean humans are replaced. We still need humans. Not least because they’re good to chat with during lunch, but also because we developed an AI tool which a human photo editor can use to be more efficient.
We don’t want to replace people because humans and bots have really different characteristics. We think it’s smart for humans and AI to work in tandem because by combining these two skill sets, we can optimize the post-production process. Take for instance, quality control. We call this quality assurance, and we use both humans and bots to do it. Yes, it can be done by a bot, but to make sure our quality base level is met every time, we pop lovely people at the very first, middle and last stages of our process. Humans are our All-Seeing Quality Assurance Eye.
Because the human brain excels at visual and auditory input interpretation, when we train the AI, living breathing beings make sure that it learns the correct information (so it can make an algorithm, which will eventually be automated). When an AI photo editor needs a helping hand, its output is routed to a human photo editor that amends any hiccups and keeps the process flowing. At the end of the line, a human makes sure the output is top-notch and signs off images before they head to the customer. It’s a self-improving loop meaning our quality assurance and quality base level bar(s) are consistently being raised. We get training data, train the AI photo editor and when it needs assistance, a human steps in. Using this, we can create better training data, re-train the AI and so on. Over time we improve, making for a higher standard of quality all round. Yippee. Humans are an integral part of what we do, and how we do it so well. AI image editing simply gives them a little pep in their step.
AI makes post-production faster. People and AI photo editors work together at Pixelz, but not in the same timeframe. Images flow between AI ‘eyes’ (creepy) and human eyes (phew) with no bottlenecks (in theory, in practice we still have some bottlenecks). Just like that, we minimize the lead time so we don’t have a human waiting while the AI is processing things.
AI makes post-production cheaper. The cost-per-hour of a human compared to a computer just isn’t the same. In fact, they’re incomparable. Still, there are some things which AI image editing can’t do as well as humans, so we use both. It is the optimisation of using AI photo editors and human beings that ensures our process is cost efficient and high quality. In-fact, 60% of our AI-made masks do not need additional work, which means 15% savings on production costs. Win.
Bots can fail. No questions asked, and they can fail very badly. But when they work, they tend to be more reliable than humans. Data from 2021-2022 shows that our customers rejected more images edited by humans compared to bots. We know what you’re thinking, will AI replace us? Are we doomed? We sure hope not. We need (and love) reliable humans and their high quality assurance. Speeding up e-commerce post-production, at Pixelz we are all about time and cost efficiency, and streamlining the process, to get products online and in shopping baskets, pronto.
How does Pixelz use AI in post-production?
Did we mention that some of our algorithms are insanely complicated? We won’t go deep into detail but here’s a brief introduction into AI photo editing and automation at Pixelz.
When we talk about automation for image editing we mean two things - automations and algorithms. Automations make things happen automatically (duh) and for those to work, we need algorithms (this is where the magic happens).
For instance, ghost mannequins. When you want a product to look like it is on a model and there is no visible mannequin, it’s not just a case of copy and paste. Images need to fit in the right way, so it really looks like the inside of that bra, shirt, dress or whatever garment you want floating around in space. We designed an algorithm that breaks down all the editing into small steps, and then automated it. Some of these steps are completed by a computer (AI and other algorithms) and some of them are done by our photo editors (IRL human beings).
First we'll use a human for what they’re good at, understanding images and making decisions, such as where to place the inset images so they look right. Then we hand over to several computer editors combining AI and other algorithms for things like masking, warping and shaping - these are all automated. Then we’ll throw the image back to a human to check that the mask, warp and shape look right to the human eye, to amend if necessary. The final quality assurance step is always humanized, where images are either rejected and tweaked manually, or approved, sealed with a bow and a kiss, and sent off to the customer. We use automation because it does the job accurately down to the pixel. A labor intensive and grueling process for others, at Pixelz 70% of the work we do is automated, and no one does it quite like us. Our AI image editor can do a ghost mannequin as well as, if not even better than a human, and always in far less time.
We’ve also got a neat thing called a trimap which breaks images down into three parts - the foreground, background, and the border (an outline for where the mask will be). Actually, we’ve got algorithms and automations for loads of time-consuming tasks like making paths (for drawing outlines around products) or color matching and changing (for garments, products and backdrops). A first ever, we’ve even designed an automation that quality-assures masking. That is, we successfully got an AI to quality assure an algorithm which it, itself, made in the first place. Meta.
While we use AI to make our algorithms, the algorithm that puts everything together is not AI. Why? Quality is our top priority and on occasion, AI image editing will have a field day. For things like merging images or adding shadow, AI can’t give us the high quality guarantee that we need in the final image. Minor details are a big deal for us and our customers, and things like a slight change of color or detail just won’t do. For this reason, we use non-AI algorithms to put everything back together again (Humpty Dumpty style).
With e-commerce's need for speed, we’ve cracked the code on how to whistle-stop through masking images - by moving it to cloud-based servers. We can scale up and down how many servers we run on at any one time relative to our workflow, keeping our processing speed consistent with a guaranteed maximum time of 5 minutes (no matter if you’ve got 20 or 20,000 images, they can all be masked in this timeframe. Cray).
Currently revamping and improving the communication between our human and AI photo editors, we’re in the process of pushing our automation rate even further. When we roll it out, we will be even more specialized at delivering overall higher quality at a faster speed. Already ahead of the competition technology wise, we are going to be even further ahead in the near future - watch this space.
Can bots be creative? AI photo editing VS traditional e-commerce photography
Define creative. If we’re talking about the ability to create an image from nothing (like random values) then, yes. If it means creating an image from a text prompt, then, sure. If it’s about being able to create text from a random input or text prompt, then again, bots are creative. But the challenge isn’t really whether or not they can do it, it is about our perception of what they do. The question is about the realism of AI image editing. How ‘real’ is what they create, to us?Traditional or AI, the nature of e-commerce photography means there is always going to be a gap between the real in-store product and what we see online. This has always been the case (it is a photo, not clothing, darling).
Traditional or AI, the nature of e-commerce photography means there is always going to be a gap between the real in-store product and what we see online. This has always been the case (it is a photo, not clothing, darling).
Rather than asking ‘does it look the same?’ it’s more about, ‘how big is the gap between the photo and the real product?’ What we need to consider is, is the gap with the real product bigger in one case or the other? If the gap is the same or smaller than with normal photography, perhaps AI is good enough? Like all of us though, AI photo editing does have limitations.
Post-production optimization is the way to go
The aim of the game is to optimize time, cost and quality. Quality comes down to two things - consistency and reliability. Looking at humans (yes, you) they're not so good for cost, time efficiency, or consistency, but they sure are reliable. With AI image editing and automation, cost, time and consistency are very good, but they're not super reliable.
If we want a future for e-commerce post-production that is high quality, and time and cost-efficient, a creative collaboration between human and AI photo editing is the way forward. Pixelz synchronizes these seamlessly, so you can get your products online ASAP. Want to give it a go? For questions and to learn more about what we do and how we can support you and your brand, feel free to get in touch.
