Autonomous Fashion Trends with Stable Diffusion
Design your wardrobe with AI
I want to buy these denim t-shirts. Not much of a fashionista, but I haven’t seen these designs before. Can someone make it?
These patterns reflect the latest consumer interest and are automatically designed by AI. Autonomous is a better word… a self-driven machine!!
How?
The machine has 3 primary components as seen below. It has no manual input. There is a deluge of data on the internet. People search, transact and leave huge data trials.
Unearth Trends: the first step is to mine these internet data trails. The machine programmatically understands what consumers are searching for, design patterns, colors, fabric, graphic styles, etc. One can in addition mine trends in music, movies & entertainment on a given day and overlay those trends to enlarge the style set.
There are tools available like decode’m (www.decodem.ai), MeetGlimpse (meetglimpse.com), Treendly (treendly.com), etc. that do this today. They autonomously track trends, synthesize insights using Natural Language Processing (NLP), generate reports and some even expose APIs to leverage in use cases such as the one we are discussing. Below are the current trends for denim t-shirts
Prompt Engineering: The second step is to use these trend keywords to generate prompts that describe what we want to see. Many large AI language models have been trained on 100s of billions of words like GPT3, GPT-J, etc. They possess a lot of knowledge and connections that can be fine-tuned for a particular task. This is accomplished via a technique called prompt engineering. You show the AI a few examples of what you will give and what you expect out of it. In this case, we tell it that we will give you a few words and you have to give us a descriptive sentence that can be used to generate an image. That’s exactly what was done. We provided the prompt to the left (figure below) and the AI is intelligent enough to understand the task. When you feed more words it gives you prompts like the ones to the right. “A denim oversized t-shirt with high collar and graphic on the chest” is an example output generated by the AI model. The technique is called “few-shot learning” as you learn from a few examples (3 in this case).
Image Generation: We have the prompts and are now ready for the magic. A prompt like “product photography of a t-shirt, denim, with half sleeve, highly indicate, professional photography” provided to a different AI model would generate the images you saw at the beginning of the article. Some amazing AI models have been released recently like Dall-e, Stable Diffusion (by Stability.ai), MidJourney, etc. that can take an image prompt and generate stunning images. These are essentially text-to-image converters. We use stable diffusion here.
We have many combinations of trends (color, fabric, other entertainment categories, etc) and hence can generate a huge number of prompts in one go. For example, when the prompt generator picks a different color (in this case red which is also trending) you get the below designs
The prompt generator uses the music, movie & entertainment trends and spots a movie Katputhli trending in India and generates the below designs
It is a machine. No egos no laziness. Prompt again and get more
It produces designs for women. You can see some different types of denim dresses generated too.
And even logo concepts. Remember we are going through a massive loop of prompts with trending consumer concepts.
Are you wondering if this was by chance? And whether it will work for other apparel? Here is a USA-focused machine that captures trends from that geography and applies them to Jackets. Note that all trends have changed from before. For example, you can see Rings of Power & House of Dragons trending. Different fabrics and colors are trending.
Perform the same steps of prompt engineering and image generation and this is what we get.
“Design to inspire” from different fabrics, styles, gender, and even futuristic-looking designs.
At one shot 100s -1000s of designs are generated. The machine in our case runs autonomously every day and the designs are unique. The frequency can be what you choose including on-demand. This can fundamentally change the ideation & design process in fashion brands & marketplaces. There will be no dearth of ideas and the designs are driven by the latest trends that consumers want.
As one generates designs the learning loop for prompt engineering can be closed. For example, we realize that the order of words matters. “t-shirt, denim” is better than “denim, t-shirt”. Using NLP techniques prompts can be engineered to better represent these learnings.
We can change the category further and extend the concept to shoes for example. Enjoy the variety of designs below.
And more designs by further prompt engineering. The words that we choose and how it has to be presented can be changed. For example, do you want the output like a product photo, a catalog, a painting, or a poster will change the output styles?
Imagine this. You have to go to an event. You run this machine, get multiple designs basis the latest trends, 3d-print the design and wear it. The day is getting here. How will designers, product manufacturers, and e-commerce players adopt to that?
And if you thought the apparel industry is ripe for disruption… enjoy these designs
Multiple sectors that leverage visual design like automotive, consumer durables, D2C & B2C merchandise, interior design, etc should be thinking machines rather than human-driven designs.
Today most design processes are human-led and are constrained in terms of variety. Manufacturing & distribution technologies are already evolving to cater to a higher variety. A good example is the fast fashion trend where multiple waves of fashion are brought to retail. The design will then increasingly become the bottleneck and expensive due to talent shortages. Using methodologies like the one demonstrated here, this can be unconstrained and consumers will relish the choices leading to innovation possibilities. Not just that consumers can be allowed to prompt and design their stuff flipping product design on its head… product companies could envision themselves as enablers of consumer-designed products. Hence given these advances in AI, companies should think about machines (AI) first rather than humans' first processes.
The future looks exciting and like any technology progress, the right balance needs to be achieved for it to sustain. The massive proliferation of designs & de-bottlenecking should not lead to the current environmental issues in fast fashion. AI-led automation begs the question of job losses. I am a believer in the human tenacity to evolve. While consumers can create and choose designs they can’t decide what design goes well with what fabric, what color combinations work well, etc. The role of designers may evolve to curation & guiding. The deluge of data is real and will similarly lead to a deluge of prompts. Curating valuable prompts is a new career (“prompt engineers”) that will emerge in these organizations. eCommerce portals will not just have products but “text prompts” & recommend new prompts instead of products. A lot is possible and the bet is that human tenacity & creativity will shape a new sustainable future leveraging AI.
Subscribe to DDIntel Here.
Join our network here: https://datadriveninvestor.com/collaborate
