Choosing One Pen-and-Ink Style Instead of Every Pen-and-Ink Style

Ace helping me decide on pen and ink style

Once I established which images I could reasonably use, I ran into a different problem: “pen and ink” is not one style.

The phrase sounds narrow until you begin searching museum collections. It can describe loose sketches, elegant contour drawings, dense cross-hatching, brush-and-ink work, editorial cartoons, botanical studies, technical diagrams, comic art, engravings, etchings, woodcuts, and drawings softened with gray wash.

Put all of that into one dataset and the result might look vaguely old, monochrome, and illustrated. That is not the same as teaching a model a controlled visual language.

The first LoRA needs a more specific target.

What I Actually Want

The goal is black-and-white fantasy storybook illustration: confident drawn contours, visible pen work, hatching and cross-hatching, readable silhouettes, expressive figures, creatures, buildings, landscapes, and complete scenes.

I want the LoRA to help illustrate original characters and settings from the Ayam Handsoom stories. That means the dataset needs to teach a reusable drawing style rather than one character, one costume, or one artist.

This also means rejecting images that are attractive but point in a different direction.

Close Is Not Always Close Enough

Engravings and etchings can resemble pen drawings because both use dark lines on light paper. But the texture of a printed plate is different from the movement of a hand-drawn line. Woodcuts introduce another kind of repeated mark. Gray wash adds broad areas of tone that may compete with the linework.

Any of those could support a good LoRA. They simply belong to different experiments.

For the first run, I am excluding gray wash, engravings, etchings, lithographs, woodcuts, and related printmaking styles. I am also excluding color illustration, photography, flat modern vector art, pages dominated by text, and scans damaged badly enough that the defects could become part of the learned style.

The purpose is not to declare those media inferior. It is to keep the first question answerable: can a small, carefully selected dataset teach a recognizable fantasy storybook pen-and-ink look?

Consistency Still Needs Variety

A style dataset can be too narrow in the wrong way.

If most of the images show a single standing person, the model may learn that composition along with the linework. If they all come from one book, it may learn recurring costumes, faces, or page layouts. If they all come from one artist, the project starts looking less like a study of shared craft and more like an imitation exercise.

The final set should therefore include several artists and a balanced range of subjects. My current target for at least fifty images includes:

• individual characters and portraits

• action and multi-figure scenes

• creatures and animals

• buildings, landscapes, and environments

• props and unusual compositions

I plan to collect more than fifty candidates and reject aggressively. Fifty is the minimum final set, not the number at which I stop looking.

Quality Is Part of Style

A perfect drawing can still be a poor training image if the scan is tiny, blurred, badly compressed, stained, warped, or covered in typography. Those defects are information too. A model does not know which parts I intended it to learn.

I will preserve every original download and process only copies. Cropping, straightening, removing empty margins, converting to grayscale, and making modest contrast corrections are reasonable. Generative restoration and AI upscaling would guess at information that was not present in the source, so they are outside the rules for this experiment.

Duplicates will also be removed. Two scans of the same illustration do not create two examples; they simply give one composition extra weight.

Why I Am Leaving Out My Own Art

Original drawings by Michele or me would be easy to document and safe to use, but I am leaving them out of the first run.

This experiment is specifically about whether an independent creator can build a coherent style LoRA from verified public-domain material. Adding our own work would change the question. It may make sense later for a custom house style or character LoRA, but keeping Run 1 clean will make the results easier to understand.

The Dataset Is the Design

The training settings will matter. The captions will matter. The base model will matter. But before any of that, the dataset defines the assignment.

I am not collecting every image that could be described as pen and ink. I am selecting images that agree about the kind of pen-and-ink world they are trying to create.

Sources consulted

The Metropolitan Museum of Art, The Printed Image in the West: History and Techniques Hugging Face, LoRA Hugging Face, Create a Dataset for Images Zeng et al., Infusion: Preventing Customized Text-to-Image Diffusion from Overfitting Somepalli et al.,

Kim, Chen, and Qiu, Learning to Customize Text-to-Image Diffusion In Diverse Context

Understanding and Mitigating Copying in Diffusion Models