Style LoRA

ELI5 (Explain Like I'm 5)

• The Big Idea: The same tiny trainable "sticky note" that can teach a model a new subject (like the LoRA Fine-Tune project's bag) can just as easily teach it a new look instead. Train it on a batch of images that all share one visual style, and it learns that style itself — well enough to apply it to things it never even saw during training.
• Analogy: Show an art student thirty drawings that were all made with a thick marker. A good student doesn't memorize those thirty specific drawings — they notice the technique (thick, bold strokes) and can apply it to something brand new you ask for. A student who instead just memorized and redrew the thirty originals has "overfit" — they learned the pictures, not the technique, which is exactly the failure mode this project is built to catch.
• Example: We train a LoRA on bold-stroke versions of digits 0–4 only, then ask the model to draw digits 5–9, which it never saw in the bold style. They come out bold anyway — proof the adapter learned the marker technique itself, not just the five training pictures.

Key Insight

A style LoRA shows that the same low-rank fine-tuning trick used for subjects works just as well for a look — train on ~30 images sharing a coherent style (a painter's brushwork, a film's color grade) and the adapter learns the style without memorizing any single picture. Success is measured by generalization: the style must transfer to prompts that never appeared in the 30 images, proving the LoRA captured how things are rendered rather than what was in the training set. This is the line between a useful style adapter and an overfit one — the former restyles anything you ask for, the latter just regurgitates its training images.

What's in this directory

File	Role
style_lora.py	bold_style (the target look), then inject the project-50 LoRA into the class-conditional base and train it on styled digits 0-4 only

The method — LoRAConv2d and inject_lora — is imported directly from the LoRA fine-tune project; nothing about the adapter changes when the target is a style instead of a subject. The base is the shared class-conditional DDPM.

# reuse the shared conditional base (also used by projects 51/52/55):
python ../51-dreambooth/train_cond_base.py --out checkpoints/cond_base.pt
python style_lora.py     # ~3 min

The generalization test, by construction

The style (a bold "marker" render — every stroke thickened) is only ever shown to the LoRA on digits 0–4. Digits 5–9 are held out of training entirely. So if the model draws a bold 5–9 at the end, the adapter cannot be replaying training images — it must have learned the rendering, decoupled from the content. That is the whole point.

Results

The target style, on the training classes (real digits 0–4, then their bold-styled targets):

The test that matters — held-out classes 5–9. Top two rows: the frozen base (its normal, thinner hand). Bottom two rows: the same classes with the style LoRA. The 5–9 come out visibly bolder even though the adapter never saw a styled 5–9 — the style generalized to unseen content:

Parameter accounting (outputs/params.txt):

trainable style-LoRA params: 39,168 (rank 4)
trained on styled digits (0, 1, 2, 3, 4), tested on (5, 6, 7, 8, 9).

Same adapter, same size as the subject LoRA — only the training target changed from "a bag" to "a way of drawing."

Things to try

• Shrink the training classes from 0-4 to just 0-1 and see when generalization to 5-9 breaks — the data-diversity floor for learning a style vs memorizing it.
• Swap bold_style for a different look (invert, emboss, a stripe overlay) and confirm the pipeline is style-agnostic.
• Stack this style LoRA with the subject LoRA at sampling time (add both residuals) — the "style + subject" composition that makes LoRAs so practical to share and mix.