fVision
fVision
an exploration of Machine Learning, Fashion and Diversity within the industry
Concept.
This idea emerged when I recalled the viral news from last October about Sarah Burton stepping down as the creative director of Alexander McQueen. She was succeeded by Irish designer Seán McGirr. This succession decision sparked backlash, as all the creative directors of these labels are now white men. Kering, the luxury fashion group, owns McQueen along with other prominent labels such as Gucci, Bottega Veneta, Saint Laurent, and Balenciaga.
I plan to explore this topic from a female perspective, highlighting some of my favourite young women fashion designers. I will speculate on how they might reinterpret collections from Kering’s historical brands if given the opportunity to showcase their visions.
I plan to use an AI model to combine the current garments created by Kering's all-white male creative directors with the styles of independent women fashion designers I admire.
Appearance:
steps
Data Scraping
AutoEncoder VAE
I decided to try the AutoEncoder VAE with a similar dataset- this are the results → not very good either after 100 epochs.
I had to have a change in plans and realise that I need to keep adding images so I can get better quality and then decide with which one I want to proceed.
I decided to combine multiple women fashion designers instead of just one and add them to the database to do a collective aesthetic and hopefully get better results.
I also transitioned from product photos to photos from the runway as its easier to keep the same format.
DCGAN - Deep Convolutional Generative Adversarial Network
I initially opted to utilise this model for training because of its capability to generate novel images based on the dataset of garments. This approach allows the model to learn intricate details and patterns from the existing designs. Moreover, I plan to leverage interpolation techniques, which will enable the model to blend and interpolate between different styles and elements present in both the current designs by male creative directors and those envisioned by the young female designers I admire.
Epoch 40:
Epoch 100:
Because the quality of the images remained absolutely horrible no matter what, I decided to resize them from 64 to 256 pixels. From there, I went through a lot of tweaking to adjust the code, which was originally built for 64-pixel images, to work with 256-pixel images. Despite my efforts, I couldn’t figure out how to make it work, so I asked for help online. Finally, I added enough layers to the generator and discriminator to achieve the desired shape, but this time the content was just noise.
Real images:
SPLICE AI
It is a model that essentially takes two images and combines them. While I would have preferred the AI to learn the styles of two different designers and generate new designs blending both styles, this was my next best option. In hopes of achieving a proper result, I began working on it.
Splice AI draws inspiration from Neural Style Transfer (NST) that represents content and an artistic style in the space of deep features encoded by a pre-trained classification CNN model.
I downloaded one image from the latest Bottega collection and another from DI PETSA, a small independent Greek fashion designer. After tweaking the code to operate independently of Google Collab and use images from my folder, I managed to generate this image:
Appearance:
I chose to adjust the transformers code provided by the original AI. Given that men's fashion is generally simpler and features fewer patterns, I found it necessary to modify the transforms. I altered the crop size, the probabilities for horizontal and colour jitter, the kernel size, and the Gaussian blur.
Structure:
Structure:
Result:
Appearance:
Appearance:
Structure:
Result:
Result:
Structure:
Appearance:
Structure:
Result:
Fake images:
Result:
Conclusion
Although I have spent a day and a half to download a dataset of 6000 images plus I don’t even know how much longer to edit them- crop, add borders, remove background, make them black and white(I didn’t even mention this last failed attempt at getting better results) - I am happy I got a decent result with the Splice AI that needed only 2 (T-W-O) images 🙃
In this process though I got even more accustomed with two AI models from class: DCGAN and AutoEncoder, and have learned about a new one, Splice AI. I understood the limits and capabilities of all of them and grew more comfortable with debugging and testing code in the pursuit of good qualitative results. I also increased my patience levels.
Based on the results I've obtained, for future projects, I'm considering taking some designs and recreating them in 3D software such as Clo3D, Cinema4D, or Unreal. This approach aims to enhance clarity and truly leverage technology to challenge the status quo of the fashion industry. By providing people with a glimpse into an alternative reality where creatives are given equal opportunities to showcase their talents, we can push boundaries and foster inclusivity in the industry.
I think the main lesson that I got from this is that, contrary to popular beliefs, its not the journey that matters but the destination and somehow I got there in the end! 🦋