FashionGAN

Motivation

The project began with a shared goal: changing how fashion designers work. Understanding the changing fashion landscape, our team aimed to innovate and tackle a common challenge: the search for new ideas and simpler design methods.

We wanted to empower designers with a modern tool that goes beyond traditional limits. We envisioned an AI-powered tool not only creating new clothing designs but also inspiring designers looking for fresh ideas.

Our project focuses on an AI-driven platform generating unique clothing visuals. By combining technologies like DragGAN and AutoEncoder, designers can access and personalize these designs to match their artistic vision.

In essence, we aim to offer designers a valuable resource that nurtures creativity, fosters experimentation, and speeds up design work. Our tool aims to be a catalyst for transformative innovation in the fashion industry by opening up limitless possibilities for designers.

Deliverables

Dataset

A folder with 8249 images from 266 fashion shows, scraped from NowFashion. The images are cleared and converted. We removed the pictures where a face wasn't recognized by our algorithm, and converted all the pictures to 256x256 pixels, as the images need to be square with a number of pixels that is a power of 2. To go with our weak ressources, we chose 256 pixels instead of a more convenient 1024 pixels.

Software

As we modified the interface of DragGAN, we've put the updated versions of all the files we modified in this sense, with explanations in the README file to how to apply these updates.

We've also added to the Github the pkl files containing the weights and configuration of the model we trained for our dataset, as well as examples of generations our model did.

We also provided the Python code we used to scrape the data from the website, as well as the code we used to do face recongintion and convert the size of the images.

Milestones and Project plan

Milestones

Milestone 1: DragGAN

• Understand how DragGAN works
• Find a dataset that would be appropriate for our utilization
• Train StyleGAN

Milestone 2: Texture swap

• Find a way to apply a texture change on an image
• Train the Swapping Autoencoder for Deep Image Manipulation
• Implement the Texture swap interface in our project

Milestone 3: User Interface

• Change DragGAN's interface to make it more intuitive for our project

Milestone 4: Deliverables

• Deliver the code on Github
• Write the wiki page
• Prepare the presentations

Project Plan

Week	Tasks	Completion
Week 3	Choose our project between the ones that were presented in class Prepare for the first presentation	✓
Weeks 4-5-6	Find some clothing datasets Read papers, understand how DragGAN works Define the project Get familiar with the libraries we want to use (selenium, opencv...)	✓
Weeks 7-8-9	Scrap the data from nowfashion website Resize images Clean dataset (remove the pictures that aren't in the same format as the other) Get familiar with StyleGAN and its functioning Find a textile dataset	✓
Week 10	Define a precise planning for the following weeks Prepare for the presentation Write the Project Plan and the Milestones on the wiki page Figure out how to implement the texture changes Begin the training of the dataset with StyleGAN	✓
Week 11	Install a Virtual Machine to use linux Finish the training on the VM Get DragGAN to work on the VM	✓
Week 12	Work on the texture changing tool Implement the texture changing interface Start writing the description of the methods on the wiki page	✓
Week 13	Adapt the DragGAN GUI to be more intuitive Finalize and clean the code to then commit it to Github Finish writing the wiki page

Methods

For our project, we're basing ourselves on the DragGAN project (https://github.com/XingangPan/DragGAN), into which we're importing a StyleGAN model that has trained our database.

Scraping

We scraped the images from the NowFashion website, containing an enormous list of fashion shows and pictures of their clothes. We scraped using the selenium library, by iterating on all shows and saving the websites. A big majority of the pictures are in the same format and have the same "frame" (the models appear in the same way in the picture).

"Normal" frame

But eventually, there are some picture completely different of the others that we wanted to erase.

"Weird" frame

The best idea we found to find these was to use a face recognizer. So, we removed the pictures where a face wasn't recognized by our algorithm, and converted all the pictures to 256x256 pixels, as the images need to be square with a number of pixels that is a power of 2. To go with our weak ressources, we chose 256 pixels instead of a more convenient 1024 pixels. All the code is available on our Github.

Copyright

The images used in this project are scraped from NowFashion illegaly. We've used these images purely and simply for learning and studying pruposes. If we had wanted to publish a paper or go further, we would have had to either contact nowfashion and find a way to legally have an access to these images, or find another source of images.

DragGAN

DragGAN is a pioneering deep learning model designed for unprecedented controllability in image synthesis. Unlike prior methods reliant on 3D models or supervised learning, DragGAN empowers users to interactively manipulate images by clicking handle and target points, moving them precisely within the image. By utilizing GAN feature spaces, this approach allows for diverse and precise spatial attribute adjustments across various object categories. The model enables efficient, real-time editing without additional networks, facilitating interactive sessions for layout iterations. Evaluated extensively across diverse datasets, DragGAN showcases superior manipulation effects, deforming images respecting underlying object structures, outperforming existing methods in both point tracking and image editing.

StyleGAN

StyleGAN is a type of generative adversarial network (GAN) developed by NVIDIA that's primarily used for generating high-quality, realistic images. It's renowned for its ability to create lifelike human faces, animals and objects. On all versions of StyleGAN that exists, we chose to work with StyleGAN2-ADA (https://github.com/NVlabs/stylegan2-ada-pytorch), based on StyleGAN2. It is less powerfull than StyleGAN3, but regarding our ressources a model based on StyleGAN2 is more appropriate. ADA stands for "adaptive discriminator augmentation", it signifies that this model is better suited for smaller datasets (for example, StyleGAN2-ADA aims to have the same results than the original model with a 30k datasets instead of 100k).

Texture Swapping

The Swapping Autoencoder is a specialized deep model designed for image manipulation, not just creating random images. It works by encoding an image into two separate parts and ensures that swapping these parts creates realistic images. One part encodes the structure while the other represents texture by capturing patterns across different areas of the image.

The goal using this was to give an additional tool to designers to modify the generations, applying the style and texture of other images on the image they are working on.

Discussion

Limitations

There were several unexpected issues during our attempt to train StyleGAN with our dataset. Without going into details, the different problems we encountered were related to :

• Wrong OS
• Use GPU
• Wrong compiler used
• Not allowed to install modules/libraries

Due to these issues, we've been able to start the training only 24 hours before the limit hour to write our wiki page. So the result we share here are unfortunately below the result we could have had with a little more time of training. Also, to have the best results in a minimum of time, we also chose to diminush the resolution of the pictures so the training would take less time. Also, when rescaling the images from the original resolution, we needed to make a choice as the original images were not square : crop them, add stripes on the sides or stretch them. We decided to start by adding stripes, aiming to test the training with these stripes and to also try the other options to compare.

Also, we had to sacrifice the implementation of texture swapping for lack of time, as problems with StyleGAN and DragGAN took longer than expected.

Results

Our results are rather unsatisfactory. Our model didn't have enough time to train, and ideally we would have also trained using images of better quality (1024x1024, for example). As a result, the images generated are of rather poor quality, as we can see from the attached images. On a positive note, the strategy of white stripes on the sides worked well. We could have continued to train our model by giving it images with white stripes on the sides and removing them before output.

Example of generations

On the DragGAN side, the same disappointment. Based on our trained model, DragGAN doesn't work very well. We can explain this by two different theories. Firstly, quite obviously, the model hasn't trained enough and doesn't at all produce satisfactory enough images for DragGAN to work. Secondly, we can also think of the low variance between the images generated. DragGAN works thanks to the diversity of its database; here, our generated images all look the same.

Github

FDH-FashionGAN