Cyberpunk City

The project aims to create a series of images in the style of a neon cyberpunk city using a generative neural network. The core concept revolves around the visual representation of a futuristic metropolis. The focus is on bright neon signs, dense urban architecture, light reflections on wet streets, and the nocturnal atmosphere.

This style possesses distinct visual characteristics. It is easily recognizable due to the contrast between cool and warm colors, high detail density, and characteristic lighting. This feature allowed for an assessment of the neural network’s capability to recognize and reproduce the artistic nuances of the style.

As source material, I used a set of images depicting a neon city landscape. The dataset includes streets, buildings, and signs. There are a total of 24 images. These images form the foundation of the visual style that the model must master.

The final output is a cohesive series rendered in a unified style. All images share a consistent neon color palette, a nocturnal atmosphere, and a dense urban structure.

A series of images, executed in a cohesive style, has been created. The entire collection is unified by a neon color palette, a nocturnal atmosphere, and dense urban structures.

The series features:

[1] street scenes with neon signs [2] city perspectives with receding roadways [3] scenes featuring light reflections on wet asphalt [4] high-saturation color images

Each image varies in its composition and object placement, yet maintains a singular stylistic vision.

The results indicate that the model has successfully grasped the key characteristics of the chosen style.

The main elements that were conveyed include:

[1] color palette: a combination of neon shades of blue, pink, and orange [2] lighting: bright light sources and contrasting shadows [3] atmosphere: nighttime, wet surfaces, reflections [4] city structure: dense architecture and narrow streets

The images possess a sense of spatial depth and dynamism. The model reproduces the characteristic visual environment of a cyberpunk city.

Some of the scenes are compositionally close to the source images—this is attributable to the small size of the dataset. The model tends to recall the structure of the scenes more than create entirely new compositions. However, the images are not exact copies; they incorporate variations in shape, lighting, and detail.

The differences between the images are evident in: camera angle, object density, lighting intensity, and distribution of color accents

The model partially generalizes the style, creating new variations based on the initial data. The resulting series showcases the overall character of the style alongside the scene variability.

Training was conducted within the Google Colab environment using the Stable Diffusion v1.5 model. The LoRA method was applied, which enables adapting the model to a specific style without complete retraining.

Key Stages:

[1] Dataset Preparation Resizing images to a square format Scaling to 512×512 pixels

[2] Captioning A unified text prompt reflecting the desired style was assigned to all images.

[3] Model Training A pre-trained Stable Diffusion model was utilized The LoRA method was implemented Training was executed for 200 steps

[4] Image Generation The model utilizes the text prompts The resulting images are saved as a series

The project utilized ChatGPT for:

idea selection technical error correction preparation of textual descriptions

Images used in the project are sourced from open repositories and are distributed under licenses that permit free use for educational purposes.