Toonkit's EURO lays out Style LoRA data prep: 20 images, split style from content

The guide's framing matters more than any single number in it. A Style LoRA is meant to transfer a shared way of drawing onto subjects the training set never contained, so EURO's first test is whether the same lines and shading show up when you generate an animal, a street or a forest — not whether the model can reproduce an image close to the training data.

That is why the data advice keeps circling back to variety. If every image is the same front-facing, same-haircut upper body, the person and composition can get bundled into the style. The guide's example set spans cel animation, thick oil painting, watercolor, semi-realistic digital paint and 3D rendering, and it says that starting with a single stated goal makes results easier to judge, rather than mixing widely different looks from the outset. The 20-image figure is presented as Toonkit's service-side guidance, not a rule that 20 is enough or required.

On captions, the guide is careful about a common shortcut: describing content and omitting style words is one design choice, but it is not a strict sorting where whatever you leave unwritten is learned as style. Whether to include style words depends on the base model, the training method and how the LoRA is invoked at generation. Its suggested settings — Step ranges, Rank 8/16/32, Learning Rate 5e-5/1e-4/2e-4 — come with the caveat that the original guide omits base model, optimizer and tool versions, so they are candidates to compare, not a portable preset. The evaluation routine it adds is deliberately plain: fix the same prompt, seed and settings, then score style, instruction-following and breakage separately.