Awesome. Okay.

Introduction

So my name is Ava Oppenheimer. I'm the founder at Trieto. I'm going to get on to more about what we are doing, but first I'm going to hit us off with a fact.

Aviation accounts for 2.5% of global emissions, carbon emissions, sorry, annually. The fashion industry contributes to around 10%.

Now, the thing is, this isn't just a statistic. This is a wake-up call.

And unfortunately, we have known about issues with sustainability for years, and despite expert warnings, we are refusing to do anything about it. And more to this, as a society, we are becoming more and more comfortable with overconsumption and normalizing overconsumption, and companies are becoming more and more comfortable with their greed.

This is an issue.

The Problem of Overconsumption

So as someone who has worked both in STEM and in fashion, I have seen firsthand how much overconsumption is a plague on the industry. So today I'm going to be showing you how meticulously curated data sets can help change sustainability and how you can potentially utilize this in your workflow.

So, between my STEM and my first job as a, or I worked in a tech company, Synthesia, I took a detour into fashion. I worked as a stylist on shoots where we would order a monstrous amount of clothes and return them.

What happens to these clothes? And that is what initiated this spark of interest in clothing and sustainability, and that is why we formed Trieto.

Waste in E-commerce

So the issue is 80% of e-commerce orders that are returned go to landfill. That does not include the amount that are incinerated or just disregarded.

And believe it or not, it is cheaper for companies to do this. And that's the issue, costing. If things don't benefit the manufacturer or the sales, at the end of the day, people don't care. And that is our biggest problem.

So that is, again, what really inspired me to buy Traito, is how we're not going to change society as a whole, but how we can change the tools that are there. So this can just be implemented into our day to day.

Our Solution at Trieto

So, what exactly are we doing? We are developing layered AI tools to generate 4D garments. That sounds extremely hefty, and yes it is.

I will get on to what that means later. But we are aiming to solve issues in waste and inefficiency in the fashion industry.

So, what makes us different? The approach that we're taking isn't CG, it isn't development, it isn't necessarily just rendering, but it is actually the input data that we are putting into these models.

The Role of Data in AI

So, what is the role of data in AI?

AI isn't inherently creative, but it is very responsive. Its creativity forms from the input that we give it.

So, if we think about it like this, fabric is the material for clothing. We could also, in the same way, say that data is the material for AI.

And in the same respect, poor quality material gives us poor quality clothing. So we can say the exact same about AI.

Poor quality data and messy data sets give us an underwhelming AI output. And we have seen this in a multitude of different tools.

Virtual Try-ons

Specifically, virtual try-ons are a big, big, big, big point in AI and fashion. But current tools rely on image data scraping, which is a very interesting topic. But the issue is, whilst it does produce good output, it doesn't produce great output.

So as an example of input and output data, I partnered with the founder of ExoLink who are utilising AI tools to create content. And this is a bit of a funny thing that we did.

But essentially we took an example of a bad data set, which unfortunately I wasn't able to get my hands on to show you, and a good data set. And we gave the same prompt to show the difference in output.

So the bad data set consisted of a variety of different iPhone lenses, including 0.5, which if you have had to play with an iPhone before, you would know that 0.5 just generally does distort the image. Bad lighting and a bad array of different angles that the photo was taken at.

So we gave the prompt, a woman walking in a rainforest, she faces the camera and smiles. This was the output we got. As you can see, this is me.

However, it's not really what I look like. This is my side profile for reference.

It doesn't get my outfit, which again, I do apologize, I don't have the original data set that we had, but I'm hoping to publish this somewhere so you can fully see the differences. And it doesn't listen to the prompt, weirdly enough.

And I think a lot of the time, people are blaming AI, it's not good enough, it's not good enough. But prompting is a big part of that data that we are feeding these models.

So we did it again using the same prompt with a much better data set, more images, more consistent lighting, so you can better understand fidelity of the data set. And this was the output we got.

It doesn't look exactly like me, but considering this was the first time using one data set, it was pretty good. And it does also listen to the prompt.

Then what we did is take all the data sets and compile them into a huge data set and use a very small prompt. Now this is to show, again, that data sets, yes, while they are a very important part of feeding AI, prompting is also a big part of the data that we are feeding it.

So, standing in a neon-lit street. They all kind of look like me, but the thing is here, we've been so vague that AI has had to almost create its own creativity. And you can see, you know, most of these images, I am very feminine, but there is one image where I'm much more masculine, almost look a bit more like a Navy cadet.

Don't know why that's happened, but it's interesting to see what AI is doing and where it's going and why it's going that way.

Synthetic vs Non-Synthetic Data

So this leads me to talk about datasets. Synthetic versus non-synthetic.

So synthetic data sets are artificial intelligence essentially. So they're generated based off computer vision and simulations.

Super cool because they're fully customizable, super large scale, very cost effective because you don't have to be in a studio shooting everything. And for the same reasons, diversity and control are huge because you can say, cool, I want to shoot 1,000 people.

How much is that going to cost in a studio? With diversity and control, you can say, cool, we're going to do 10,000 people for a fraction of the cost and also get a better result because the data is more diverse.

However, synthetic data really lacks realism and there is an extreme model bias. If the data that you are feeding the synthetic data with isn't good enough, your output is going to be bad, similarly to how I just showed you in the example with the generated image.

And the initial effort is huge. You need extremely sophisticated generation models to actually achieve good synthetic data.

Nonetheless, I mean, we're all here now. AI is a booming topic, and more and more people are learning about it, so we don't need to worry about experts in that field.

Now, moving on to synthetic data.

Pros. High authenticity, reliability, and no assumption. I'm going to bracket these all into one key point here, is that there is no assumption.

You are training a data set on real life, real world scenarios. So for example, there's something called volumetric capture. And if you are not familiar, that is essentially 360 degrees of body scanning. Super, super cool.

With this, you can see in a real world scenario how things are moving, how a person is acting. However, processing costs a fortune and getting people to actually be part of these data sets is a bit tricky considering the current climate of AI. Nonetheless, the payoff is extremely, extremely high.

I'm not showing this to say pick one or the other. That is absolutely not it. But hopefully you can see that some of the pros in synthetic data match up to some of the cons in the real world data. And this is to show that we can actually utilize these together to create really cool outcomes.

And that is what we are doing at Trieto. I actually decided I'd ask ChatGPT what it thought of these different types of datasets, and this was the output we got.

Synthetic data and real data are not competitors. They're collaborators. When used together, they form a powerful toolkit to train AI models that are both realistic and adaptable.

And as mentioned, that's exactly what we are doing at Trilito. We are trying to get realistic and adaptable models to train AI on.

The Bigger Picture

Now I'm going to briefly discuss the bigger picture.

So, as I mentioned, the fashion industry contributes to 10% of global carbon emissions.

Everyone in this room has clothes on, and I'm sure everyone in this room has ordered clothes at least once in their life. 1I know my mum, for example, when something goes on sale, she's like, oh, top for five quid.

Am I small? Am I a medium? Am I a large? I'll order all, and I'll return some.

And that is a habit that we are normalizing in society, and we are not thinking about the results of our actions. And quite frankly, I'm not convinced too many people care, which is really, really sad.

So, what can we do to change the way we are in the world?

Improving Virtual Try-ons

We want to improve virtual try-ons at Trieto with better data so we can reduce returns significantly.

Imagine designing a garment virtually. It fits perfectly, and you can see that on unique body types, all before a single piece of fabric is cut. This way, creatives are still creating, companies are still profiting, and the world is still surviving.

And I think that's the main thing. There's no need for us to compromise anything here. But we can utilise AI for sustainability in this way.

Enhancing Creativity with AI

It is important to note that AI isn't here to replace creativity. It's here to enhance it.

By focusing on how we curate and personalize datasets, we can unlock AI's full potential. And if you're working with AI in your workflows, I ask you, what are you inputting the data with and what is the output you are getting?

Conclusion

And that kind of leads me to more of a Q&A. I apologize, I don't have a demo. But yeah, I just want to open up. Thank you very much for having me and thank you, Josh.

Is it more on the design side, so the companies will design clothes that are less likely to be returned, or is it more on the buying side, you're less likely to buy things that you'll have to return, or both? Yeah, great question. So in essence, what we are as a research company, we are researching multitude of different ways that we can, again, incorporate AI into sustainability. So it is actually for both.

What we are doing is trying to create a very realistic virtual try on tool. So you can see, again, without trying on real clothes, what it would actually look like on your body. Further to that, there are really cool tools such as Clo3D, where as a designer, you can actually use 2D manuscripts, which are called tech packs, and see what they would look like in a 3D asset. And we are going to do something similar, but just implementing slightly different technology so you can see real life physics. And then again, so for designers and for bigger manufacturers, I hope that answers your question.

And what's the biggest problem you're getting with getting the real data? Is it people just don't want to give away their body image or what? Yeah, no, that's a great question. So as I briefly mentioned, volumetric capture, that is one of the ways that we are getting this real-world data. So we are combining synthetic with real-world data.

A big issue is getting people on board with things. I think people are very scared about deep faking, and if they are part of these AI data sets, what that means for them, longevity-wise, if they're selling their souls away to AI. So I think that's the biggest hurdle that we're dealing with at the moment, on top of the cost of processing. It's a really big thing at the moment. And you don't want to try and create a sustainability tool and have intense amounts of compute power, because obviously we know what that's doing for the planet as well. So, yeah.

Have you got an idea of what other people are doing in this field? Yeah, that's a great question. There's some really cool tools out at the moment, specifically with virtual try-ons. A lot of people are doing 3D, which is really cool. Obviously, we've all seen 3D before.

But I think the issue is we've got into a habit of talking about 3D as it is 2D. And this is the thing I've seen a lot is we're essentially trying to generalize depth of field and make up what physics is and assume what physics is. What we are doing is 4D. So we are integrating time into our training models.

And by that, we are actually saying, OK, what is at the back of my jumper that you can't see that's going on? And how is that affecting the entirety of the cloth, for example? And the good thing with volumetric capture is you can actually... in video, see what is going on, and then train that into your models.

So I hope that helps, but that is our unique approach, is going for 4D over 3D. But yeah, if that's all, thank you guys very much.