I enjoy making quick sketches—sketches. And I’ve decided that I want the neural network to be able to draw in this style as well. Furthermore, unfortunately, I am completely unable to draw people, and even in a genre that forgives a lot, I cannot get them right. If I can’t do it, at least let the neural network do it.
Learning Process
To specifically train the neural network, I created 30 quick sketches of random images from my gallery.You can view them here
Self-learning proceeded according to the example provided to us. During the training, 1,000 iterations were performed, with one intermediate result after 500 training steps.
Program Outcomes
The diffusers library was used to generate the images, integrated with a trained LoRA adapter.
Before moving on to what is most interesting to me—generating human images—I decided to see how the training went for generating the images in the training dataset. I set the classifier’s scale (guidance_scale) to 5 to give the neural network more freedom, and for the image generation codes that were not in the dataset, I raised this parameter to 7.
My sketch/AI result based on the prompt «duck swimming in a pond, a drawing in the style of son_style»
Among all the training drawings, there was only one duck, so the neural network reproduced it almost perfectly, but there were several cats, and I think that is noticeable in the result.prompt: cute kitty, a drawing in the style of son_style
I gradually drifted away from the specifics of the dataset, so I then asked to generate a house in the Constructivism style and a bouquet of roses. Among my sketches were buildings and flowers, but not these specific ones.
a building in the constructivist style / a bouquet of roses in a glass vase a drawing in the son_style
Finally, I asked it to generate images of people, and I must say, the results pleased me
A girl in a dress and a man are standing next to each other in full height, a drawing in the style of son_style
A man at a table in a café / a street musician playing a guitar A drawing in the style of son_style
It seems to me that, based on the images shown above, we can conclude that the instruction was successful.
Resulting Series
As a resulting series, I wanted to generate a short comic. I based the plot on an anecdote that DeepSeek told me upon request.
— What is the name of a designer who creates a mockup on the first try and the client approves it without revisions?
— A figment of imagination.
To resolve library version conflicts, configure the environment for compatibility, and assist in idea generation and prompt text formulation, the neural network DeepSeek was utilized.
