The idea of creating clothes has haunted me for quite a long time – from the age of 10. However, I did not have the resources and understanding where to start.
It began to take shape in the kitchen at the beginning of February this year. We were sitting with my wife and friends and by chance it came up that one of the clothing manufacturing companies was announcing a hoodie design contest on its Instagram page.
At the time, I was slowly playing around with MidJourney, trying to create coloring pages for adults and a book for children. I did successfully create the book using OpenAI, and published on Amazon. I got access to OpenAI in August 2021, when I was still living and working with my family in Malaysia.
In the same evening, I created several designs and my friends reacted quite enthusiastically to them. I saw potential in the designs and started generating them in batches, trying to figure out how to get better results. MidJourney exceeded all expectations: even a few words of prompt (prompt) produced a lot of unusual designs that could be spun into several model collections. It was something never seen before, I felt that I was standing on the edge of an abyss – an endless space for inspiration.
I submitted a dozen designs to the contest, but I was not chosen as the winner because, according to the clothing manufacturer, it was not their style.
I’m usually driven by two forces when developing ideas: irritation with clunky analogues and the lack of a better version of an idea in the world. They were created in the same way FinGears, the prototype of which I once offered to the team TimeForMachine.
On the other hand, I am driven by the desire to automate manual work. Of course, in order to automate something, it must be systematized. So I started experimenting with different designs, trying to invent something that didn’t exist yet.
A small fraction of what I did in one day:
My MidJourney usage to date:
Somewhat systematized designs using a graphical database (if you can call Obsidian that).
This systematization relatively clearly stores the possible designs and their connection to the prompt for MidJourney.
Also, this graph can be rotated here: https://dmytrove.github.io/LatentApparelSpace/
In the desire to save money and find better ways of generating designs, I went through a dozen generative AI, cloud and local: DALL-E, Leonardo, Stable Diffusion, Fabire, etc. Everything was not the same for various reasons, mostly the complexity and quality of the images. So, I took the MidJourney subscription for granted.
As I gained experience, I began to look for ways to make a profit from it. Opened a shop on Etsy for the eighth time. As a starting point, I chose the Printify (print-on-demand) service, but I quickly realized that I could not have normal control over the quality of the clothes, and put it off for a while.
Began to look for clothing manufacturers in Ukraine. For the most part, they didn’t have the capacity or the skills to produce clothing with an all over background. I started working with two of them, but I didn’t even get to the prototypes.
At the same time, I was exploring the possibility of creating skinnies based on MidJourney images and increasing their resolution as much as possible. This is from the coloring days, I had a subscription to Adobe Creative Cloud, but it wasn’t enough. Tried various upscaling options: Let’s Enchanse, Upscaly, Adobe products, various models available on Replicate and HuggingFace. First of all, they did not like the presence of artifacts that should not be there. For now, I settled on Topaz Photo AI, which turned out to be even better for my needs than their own Topaz Gigapizel.
Deploying the images on the pattern turned out to be the most difficult task so far. MidJourney creates complex solid designs, with distortions to match, but how do you fix them?
Several approaches have been tested: reverse warping, assuming that the garment is a cylinder. I used various Python libraries for image manipulation, but they had to take into account the joining of front and back parts, sleeves. There is currently no out-of-the-box solution, but I’m pretty sure there will be a neural network-based one that can reverse engineer a 3D image into a 2D pattern. I tried Blender with drawing a projection on a 3D model. In the end I came to Adobe Substance 3D, which allows you to do what Blender does, but in better quality and with its relatively complete ecosystem.
Meanwhile, MidJourney has been releasing new tools and improving its models. I started with about version 4, currently the latest is 5.2. The tools made it possible to do everything Adobe Photoshop did much worse, even with Firefly’s new generative tool: converting an image to prompt, panning, inpainting to redraw failed parts, zooming 4x! The appearance of these tools significantly reduced the cost of the flight of thought – the embodiment of an idea in the background.
The image of the front and back of the hoodie (of course, one of the hoods should be looking at the viewer, but it turned out that way, but it is not critical).
And this is the background, which is used to overlay patterns. It was created by removing extra parts of the hoodie: hood, sleeves, etc.
Substance 3D Painter, where another design is already laid out.
Another:
Meanwhile, I realized that I don’t have enough strength to draw designs in 3D, so I started looking for a designer through my Facebook friends. There, I was advised by Kyiv clothing manufacturers who have been working on the market for a very long time and relatively successfully. I showed my portfolio, we previously agreed on the interest in cooperation. Based on the patterns provided by them, I created a layout, according to which they made a print and sewed a t-shirt – my first physical result, which finally appeared 9 months after the start of the project. The long-awaited appearance of this product gave a second wind and reinforced the belief that everything will work out.
Also, the patterns I received from the manufacturers, I stitched into a 3D model in CLO 3D – a lesser known relative of Marvelous Designer. It was this 3D model and its UV mapping that I fed Adobe Substance 3D Painter. And UV mapping – these are patterns that can be laid out on templates and sent for printing-cutting-stitching.
Looking ahead, I planned the possible formats of my store. However, I always couldn’t answer the question exactly what people like about what I do – it could be in demand. The solution seems to be obvious – we need to improve marketing analytics. But how? Ask people to take a survey where they can swipe designs left and right like Tinder? As if there were no ready-made solutions, he began to write his own. Because Google/MICrosoft Forms have their own complexities and drawbacks in creating surveys based on images. So are dozens of other survey platforms: Surveysparrow, SurveyMonkey, TypeForm … each one either has its own flaws or is too expensive if you expect to collect hundreds of votes. In desperation, I wondered if any of them could support Ukrainian Business by providing free access, if only temporarily. No one. Only some little-known Conjointly offered Ultimate access for 3 years for Ukrainians. While I was writing them an access request, it turned out that they allow you to create surveys for free (!) and have a bunch more analytics tools. It was a jackpot: in 3 hours I created a bilingual, responsive, image-based survey. Over the next day, I had results from 70 respondents, thanks to my friends who distributed the survey invitations.
Survey page:
From a statistical point of view, this was not a very representative sample. I’ve been using designs since February in the survey. However, in my head, she put everything in its place. Not everything that I thought was eye-catching turned out to be so.
It was that incredible feeling when you feel and understand your audience. I will remember the result of this analysis forever and will use this tool more than once. Also, the results of this survey allowed me to cross the point of uncertainty with the manufacturer and prove the seriousness of my intentions for production.
So, the whole process turned out to be quite difficult. Therefore, the latest clothes have not yet begun to be sewn in all the basements. However, my intuition tells me that in 5 years this process will be automated enough that for $29.99 anyone can generate 10 t-shirt designs and open an Etsy shop with them. And for $39.99 – already 100 T-shirts, and in the case of annual payment, a 20% discount. Yes, I really do not like the Pricing page, and especially when there is no price on it, but there is a contact form (which allows for entering only corporate mail) so that a manager can contact you (“I dropped the price in my inbox”).
The general process of preparing new clothing designs:
Currently, I am going to sew “prototypes” of hoodies at cost price at my own expense, in order to make photo-video material for the store and social networks based on them, to make sure of the quality.
And with them – to look for investments, offer partnerships, promote themselves, create a physical product, make a decision that can automate everything (by the way, NVidia has been for some time is working over it, and Meshy AI is already pretty good at it copes):
To make a unique design like two fingers, so that they are not torn, I am slowly thinking about a two-factor system for verifying the authenticity of physical products. Even with ControlNet QR codes, you can make cool ones, and with them – t-shirts.
At least for myself, I formulated the main principles of the project:
- Special Attractiveness: Each product is imbued with a unique essence, designed to stand out as a statement of its own.
- Refined Accuracy: With a focus on detail, we ensure that each sample is presented gracefully in crisp, high definition.
- Holistic Design: Our creations are a testament to organic inspiration, where every element is thoughtfully integrated to create a cohesive visual experience.