Training a Generative Neural Network in the Style of Sasha Robinson

I have always been fascinated by the world of fashion, which is why I decided to delve deeper into the work of the distinguished fashion illustrator Blair Brightenstein. Her pieces are imbued with a unique message: the faces and proportions within them may stray far from the classic ideal—and that is where their strength lies. She demonstrates that imperfection is not merely permissible; it can become the foundation for a true masterpiece. This idea resonates deeply with me, as it echoes my own perception of beauty and self-expression.

Blair Brightenstein’s Works

I decided to experiment and created three image series.

In the first series, I focused on depicting girls from fashion magazine covers. In my prompts, I specified various parameters: hair color and length, eye color. The result was images with abstract faces and unconventional proportions, which echoes the style of Blair Brightenstein. The color palette was also maintained in a consistent manner, lending cohesion to the series. However, there were some shortcomings: artifacts appeared in some images, and the text in the titles was inaccurate or illegible.

Prompts from left to right:

— photo collage in BREITENSTEIN style, blonde girl with blue eyes — photo collage in BREITENSTEIN style, a painting of one woman with brown hair and brown eyes — photo collage in BREITENSTEIN style, a painting of one woman with bright hair — photo collage in BREITENSTEIN style, a painting of one woman with pink hair — photo collage in BREITENSTEIN style, a painting of one woman with short blonde hair

In the second series, I decided to depict girls wearing hats from famous brands such as Chanel, Prada, and others. The result pleasantly surprised me: the neural network not only conveyed the brand names but also visually reproduced their recognizable style—their characteristic colors, shapes, and even patterns. This added depth and authenticity to each piece. As in the first series, there are small artifacts in the images, but in this context, they are not perceived as flaws. On the contrary, they emphasize that very imperfection which lies at the heart of Blair Brightstein’s aesthetic, making the illustrations vivid and expressive.

Prompts from left to right:

1. photo collage in BREITENSTEIN style, a painting of one woman wearing a Prada hat
2. photo collage in BREITENSTEIN style, a painting of one woman wearing a Louis Vuitton hat
3. photo collage in BREITENSTEIN style, a painting of one woman wearing a Gucci hat
4. photo collage in BREITENSTEIN style, a painting of one woman wearing a Celine hat
5. photo collage in BREITENSTEIN style, a painting of one woman playing tennis

Finally, I created a small series of images featuring girls in various locations—on a tennis court, in the street, and on the beach. This series turned out to be the most challenging because the high level of detail in the backgrounds and environment caused numerous artifacts to appear. However, by experimenting with the prompts, I managed to minimize these defects. As a result, the images came very close to the style of Blair Brightenstein—not only due to the characteristic color palette but also because of the girls' expressive poses, the meticulously developed backgrounds, and attention to fine details, which makes the series particularly expressive and stylistically cohesive.

Prompts from left to right:

1. photo collage in BREITENSTEIN style, a painting of a woman playing tennis
2. photo collage in BREITENSTEIN style, a painting of a woman walking down a fashion street
3. photo collage in BREITENSTEIN style, a painting of a woman swimming in a pool

First, I selected the necessary images—there were about 30. Then, I uploaded all the pictures to Google Colab and connected the required libraries for further work.

Image Link https://disk.yandex.ru/d/cv_6V_JoCi6EPg

After that, I used BLIP to generate captions for the images and inserted the Hugging Face token. Then, I set all the necessary parameters for generation to get the desired results.

Next, I saved the model and began working on image generations.

Model Link https://huggingface.co/evapashaeva/breitenstein_style_LoRA

Code Notebook Link https://colab.research.google.com/drive/1YY85U_HhNdqNCGnBkGd-kQJVwMSwvhI1?usp=sharing

I really enjoyed working in the style of my favorite artist, and I was pleasantly surprised by how accurately the neural network managed to reproduce her style. Despite minor flaws, the neural network performed the task very well, and the images were as close to the original as possible. It’s also worth noting that drawing people is the most challenging aspect for the neural network, as artifacts often appear in those cases. However, I managed to overcome these issues and achieve a satisfactory result.