Project Concept
I became curious how a neural network could replicate the style of my illustrations for a legal support service for galleries and artists, which was created by students from the Faculty of Law at HSE. I had drawn these cats for the corporate identity of one project. The project was discontinued, leaving 14 illustrations behind. A corporate identity requires many characters in various situations, and I drew each pose by hand. Therefore, the task is to train a neural network on this style to see how closely it can generate illustrations that could be used in the project’s corporate identity—and what new elements generation might bring to the project’s style.
Image set for training
The dataset contains 14 of my square-format illustrations, sized at 1500×1500 pixels, set against a white background. I trained a LoRA for Stable Diffusion XL using my drawings and generated a series of nine cats in new poses.
Illustrative Characteristics:
— Flat color fills without gradients or shadows — Black outline of an uneven, «organic» line — Fine hatching on fur, stripes on the forehead and back — Green eyes with vertical pupils, pink nose, long, thin whiskers
Results
«drawing in MIHAILOVA style, a red cat asleep on an open laptop»
«drawing in MIHAILOVA style, a red cat carrying an envelope in its teeth»
«drawing in MIHAILOVA style, a red cat jumping over a cardboard box»
«a drawing in MIHAILOVA style—a red cat sitting on a stack of folders»
drawing in MIHAILOVA style, a red cat licking its paw
«drawing in MIHAILOVA style, a red cat peeking out of a desk drawer»
Comments
The neural network managed to replicate the overall style, drawing technique, and details—color fills, stroke character, and eyes. However, errors persisted, such as extra paws, tails, and imperfect shapes. Often, the neural network failed to separate two cats, merging them into a single entity. Therefore, several more generations were attempted to create illustrations that were more accurate and presentable.
Result 2
«drawing in MIHAILOVA style, a red cat asleep on an open laptop»
«a drawing in MIHAILOVA style: a red cat asleep on an open laptop»
«a drawing in the style of MIHAILOVA: a red cat carrying an envelope in its teeth»
«drawing in MIHAILOVA style, a red cat jumping over a cardboard box»
«drawing in MIHAILOVA style, a red cat sitting on a stack of folders»
«drawing in MIHAILOVA style, a red cat peeking out of a desk drawer»
«drawing in MIHAILOVA style, an orange cat lying on its back with paws up»
The Learning Process
— I uploaded a dataset consisting of my original illustrations. — I captioned each image using the BLIP model and prepended the caption with «drawing in MIHAILOVA style.» The last name functions as a style tag, allowing the model to recognize my artistic manner. Since the BLIP model labeled the main cat «red cat, ” I also used „a red cat“ in the prompts for the series. — I trained a LoRA using the DreamBooth method on top of the Stable Diffusion XL base 1.0. — Training parameters. Resolution: 512, 6 images per step. Learning rate: 1e-4, 500 steps, saving at steps 250 and 500. To fit the training onto the free T4 GPU, I enabled fp16, gradient checkpointing, and 8-bit Adam.
