Consistency is key
Choose reference images that show a clear and repeatable visual style—whether that’s in brushstroke texture, lighting, color palette, composition, or character design. The more consistent the references, the easier it is for the LoRA to “learn” the defining traits. Consider using 1:1 aspect ratio for better alignment in training.
Avoid mixing drastically different styles, moods or formats.
High visual quality is important
Use high resolution images that are sharp and well-lit. This ensures that the fine details that define the style, like texture, shading and contrast, are all captured when you start to train the LoRA. Pencil requires a minimum of 1024x1024px and supports jpg, png and webp formats.
Capture core elements of the style
A strong set of references should reflect the key elements that distinguish the style. This might include:
Colour usage (e.g. muted vs. saturated)
Line or brushwork (e.g. fine lines or bold strokes)
Composition patterns (e.g. centered portraits vs. dynamic diagonal or varied)
Subject treatment (e.g. photorealistic vs. stylised artistic)
Consider it like curating a 'style signature' across your dataset.
Keep the subject matter focused
Try to use references where the subject is similar in type and framing—for instance, portraits, landscapes, or object studies—rather than mixing too many categories. Avoid unnecessary elements like watermarks, text, logos (unless intentional) and irrelevant background noise.
The more focused the subject, the more clearly the LoRA will learn the stylistic attributes rather than being distracted by content variation.
Use a minimum of 10-30 strong examples
While LoRA training is efficient and can work with small datasets, aim for at least 10–30 high-quality, consistent images.
Small datasets (10-15 images) create very rigid models, perfect for preserving brand elements
Larger datasets (50+ images) allow for more flexible and creative outputs while maintaining cohesion to the overall style