<h1 id="luis-royo-style-sdxl-style-lora">Luis Royo Style — SDXL Style LoRA</h1>Trigger: <code>Royo_style</code>A style LoRA trained as a tribute to the fantasy and science-fiction illustration work of Luis Royo.<h3 id="quick-start">Quick Start</h3>Add:<code>Royo_style</code>to your prompt and describe the image you want.This LoRA was deliberately trained so that the trigger carries much of the visual treatment. You do not need a long list of style descriptors to make it work.Many of the showcase images were generated from extremely sparse prompts such as:<code>Royo_style, a man and his dog</code><code>Royo_style, a warrior and his war horse</code><code>Royo_style, a priestess and her flesh golem</code>The generation metadata is included with the showcase images so you can inspect the actual prompts and settings.<h3 id="what-it-does-well">What it does well</h3>The LoRA produces a strongly illustrative fantasy / science-fiction treatment with particularly good handling of:<ul><li>Figures and faces — expressive portraiture, full figures, male and female subjects, and characters of differing ages.</li><li>Fantasy and creature design — monsters, demons, golems, animals, hybrid creatures and human/non-human compositions.</li><li>Materials and surfaces — metal, leather, fabric, fur, skin, wood, glass, wet surfaces, rust, stone and reflective materials remain visually distinct.</li><li>Dense scenes — workshops, equipment, tools, machinery, multiple characters and environmental clutter can retain a coherent illustrative treatment.</li><li>Mechanical and industrial subjects — vehicles, diving equipment, armor, machinery and other hard-surface objects work surprisingly well.</li><li>Environments and scale — the style transfers beyond portraits into landscapes, architecture, atmospheric scenes and large fantasy environments.</li><li>Sparse prompting — simple subject-and-action prompts often work exceptionally well. The LoRA does not require you to describe its own style back to it.</li></ul>It can produce both restrained and highly saturated results depending on subject, checkpoint and prompt.<h3 id="suggested-prompting">Suggested prompting</h3>Start simple.<code>Royo_style, an old violin maker teaching his apprentice</code><code>Royo_style, a deep sea diver working on the deck of a rusted salvage ship during a storm</code><code>Royo_style, a sorceress standing beside an ancient stone guardian</code>Then add content when you need more control: subject, action, environment, clothing, objects, composition, weather, etc.I recommend trying this before adding generic quality/style terms. One of the interesting characteristics of the LoRA is how much it contributes on its own.LoRA strength: Start around 1.0 and adjust to taste. <h3 id="mature-content">Mature content</h3>The source artist's body of work includes adult themes and nudity, and the training material/style representation reflects that broader artistic vocabulary. This LoRA is capable of mature and nude imagery, but it is not exclusively an NSFW LoRA.It is equally usable for clothed characters, creatures, portraits, landscapes, machinery, animals, fantasy scenes and science-fiction illustration.<hr /><h2 id="about-this-lora">About this LoRA</h2>This model is a fan-made study and tribute, created out of admiration for Luis Royo's work. It is not affiliated with, endorsed by, or produced by Luis Royo.The goal was not simply to make a model that generates stereotypical “Royo subjects.” The training experiment was aimed at learning a transferable visual treatment that could be applied to subjects both inside and well outside the expected source domain.That is why some of the showcase images deliberately use unusual subjects.A violin workshop, a diver on a working ship, an ordinary man with his dog, machinery, horses, landscapes and other non-obvious prompts are included to demonstrate what the LoRA contributes when the prompt itself isn't doing the stylistic work.<h2 id="training-captioning-philosophy">Training / Captioning Philosophy</h2>This LoRA was trained using a deliberately semantic captioning approach.Rather than filling the training captions with descriptions of the artist's visual style, captions primarily described what was actually depicted:trigger → subject → action/relationship → environment → composition/visibility → important objectsFor example, captions describe things such as a woman standing beside a creature, a figure seated in an environment, visible clothing or equipment, or important background elements.They intentionally avoid repeatedly teaching terms such as:beautiful linework, painterly shading, fantasy illustration, dramatic colors, detailed rendering, etc.The reasoning is simple:If the caption explains the style, the text prompt can become responsible for producing it. If the caption explains the content while the trigger identifies the training set, the LoRA has to learn what remains unexplained.In this experiment, what remains unexplained is largely the visual treatment we are trying to capture.This also makes sparse prompts useful as a test. If:<code>Royo_style, a man and his dog</code>produces a strongly characteristic illustration, then the prompt clearly isn't supplying very much of that appearance.<h3 id="training-goal">Training goal</h3>The objective was therefore style transfer rather than subject memorization.The showcase intentionally mixes familiar fantasy imagery with subjects unlike the training domain to help expose the distinction.Results will naturally vary with checkpoint, sampler, LoRA weight and prompt. Full generation metadata is attached to the sample images wherever possible so the examples can be reproduced and evaluated rather than treated as mystery showcase renders.

Files with SHA256: 5815bb90c2533b58d627818d3e93b3cc8daa7d8e62748012b33b85a8ca092c32