Training a style model on your own archive, and arguing for your authorship of the output
The situation
Advanced studio students have 200 drawings in a consistent hand by their third year, and the question they keep asking — usually anxiously — is whether a model trained on their work is a tool, a collaborator, or a thief of their own labour. Handing them a reading on AI ethics does not answer it. Training the thing does.
Steps
-
Build and photograph the dataset before deciding what the style is
Copy stand or scanner, Lightroom for batch normalisation
Assemble 40–100 works shot under consistent light, cropped and normalised. The selection is the artistic act here — what you include defines what the model will think you are.
What you only learn by doing it: Inconsistent capture kills student models, not dataset size. If half the drawings are shot under a warm desk lamp, the model learns “warm yellow cast” as part of your style. Also: students include their best work. Include the most characteristic work instead.
-
Caption deliberately, with a unique trigger token
Dataset tag managers, or an auto-tagger for a first pass
Every image gets structural tags plus one invented nonsense trigger word unique to the student. The tags you write become things the model can vary; the qualities you do not name get absorbed into the trigger word.
What you only learn by doing it: Students get this backwards every time. Tag what you want to remain controllable and deliberately leave untagged the qualities that constitute your style. Tag “loose ink line” and the model treats it as optional; never name it and it fuses into the trigger. Pick a trigger that is genuinely not a word.
-
Train three times and treat the failed runs as the data
Kohya_ss locally, or Replicate/RunDiffusion in the cloud
Run training at least three times at different learning rates and step counts, keeping every checkpoint. Students compare an undertrained checkpoint against an overtrained one and document where the usable band sits.
What you only learn by doing it: Always do a small-subset trial first. Students who go straight to the full dataset wait three hours to discover a captioning mistake. And recognise overfitting by what it looks like: if you can identify which specific training drawing an output is regurgitating, you have gone too far — and that output also has the weakest authorship claim.
-
Probe the model with things you have never drawn
ComfyUI or Automatic1111, at varying weights
Prompt for subjects entirely absent from the training data and see whether the style transfers. This is the diagnostic that tells you whether you trained a style or just a lossy archive.
What you only learn by doing it: Sweep the weight from 0.3 to 1.2 on the same seed and prompt, and put the strip on the wall. Students assume 1.0 is correct; the interesting band is usually 0.6–0.85. The strip is also the clearest visual explanation of what a style model actually is.
-
Make finished work where the model is one stage, not the output
Photoshop, Procreate, or the student's actual physical media
Generated material must be substantially transformed, recomposed, or returned to physical media. This is the pedagogical heart and also the legal posture — copyright protection requires sufficient human control over expressive elements, and prompting alone is not enough.
What you only learn by doing it: Require the raw generation and the finished piece side by side in the final crit. The gap between them is the student's authorship made visible and measurable — and students who cannot show a meaningful gap discover it in front of the room rather than in a grade comment.
-
Write the authorship statement and the dataset manifest together
A one-page statement; a manifest listing every source work
The statement names what the model was trained on, who made every image in it, what the student contributed at each stage, and what claim they are making. Students may argue a different position from their instructor's, but they have to argue it.
What you only learn by doing it: Make the manifest list provenance per image. The instant a student includes a collaborator's drawing or something pulled from Pinterest “just for texture,” the whole authorship argument changes — and they will not notice until forced to write the line.
Where this breaks down
The infrastructure breaks first. Training tooling on a college Windows image is a CUDA-and-Python version fight, IT frequently blocks the installs, and cloud training requires a credit card. Pilot the whole chain yourself a week before assigning it.
The dataset breaks second. If a student's archive contains work from another class with shared source material, collaborative pieces, or photographs of other people, they cannot cleanly claim the training set is theirs — and a model trained on someone else's images does not become theirs by being trained.
Base-model provenance is unresolved and you should say so aloud rather than pretending a clean dataset launders it. The student's model sits on top of one trained on scraped data currently in litigation.
There is a quieter risk: a student who trains a model on their own hand at nineteen may stop developing that hand, because the model makes repeating the existing style frictionless. Build in a requirement that the final work move somewhere the model cannot follow.
Provenance: the technical pipeline — dataset refinement, structural tagging plus a nonsense trigger token, a small-subset trial run before the full set, iterative checkpoints — is documented in Melanie Beisswenger, “LoRA Model Training for Stylized AI Generated Imagery,” 12th International Conference on Illustration and Animation, 2025, including her joint-authorship position. The “your own archive” adaptation and the classroom scaffolding are ours.