Concept
Illustrated covers of such ancient catalogs featuring seeds and crops—this is the stylistic approach I chose for training the neural network. These illustrations possess distinctly established artistic traits: a rich yet contained color palette, and uniform subject matter achieving infinite variability through the sheer number of cultures depicted.
CHANGE Project Goal — to train the generative model Stable Diffusion XL to reproduce the visual style of vintage botanical illustrations and then use it to create new, non-existent floral images in the same artistic language. A crucial aspect of the project was testing the model’s ability to adopt not specific images, but rather stable stylistic characteristics: composition, color palette, texture, and the overall character of an antique printed illustration.
The goal of the project was to train Stable Diffusion XL on the style of vintage botanical illustrations and subsequently create new floral illustrations within that same language. A vital part of this was determining if the model could internalize not individual images, but consistent stylistic attributes: composition, palette, texture, and the general mood of antique printed illustration.
Images are sourced from the public archives of internet_archive.org
Examples of illustrated covers used for instruction
Dataset
To prepare the dataset, I processed 33 images to bring them into a uniform format: a 1:1 aspect ratio and a size of 512×512 pixels.
Some illustrations also needed to be stripped of text. To do this, I used a generative neural network. Here is an example prompt:
Resulting Images
