<h1 id="version-3-update-semantic-captioning-edition">Version 3 Update – Semantic Captioning Edition</h1>Trigger: MManara_style This release is a complete retraining using a revised semantic captioning methodology developed through extensive probing and experimentation after the release of Version 2.Version 2 was trained using a more traditional descriptive captioning approach and remains available for anyone who prefers its behaviour. It served as the foundation for many of the observations that led to this new version.<h2 id="what's-different">What's different?</h2>Rather than describing every visible aspect of an image—including artistic style, rendering techniques, and character appearance—Version 3 attempts to assign semantic responsibility between the base checkpoint and the LoRA trigger.In practice, this means:<ul><li>Stable artistic style is largely left for the trigger to learn.</li><li>Character identity is left for the trigger rather than being described in captions.</li><li>Captions focus on concepts the base checkpoint already understands, such as pose, clothing, composition, camera angle, environment, and visible actions.</li></ul>The goal is not to describe the artwork more completely, but to describe only the concepts that should remain the checkpoint's responsibility, allowing the trigger to focus on what actually makes Milo Manara's work distinctive.<h2 id="why-retrain">Why retrain?</h2>During the development of several later LoRAs, I began probing checkpoints with sparse prompts to understand what they already "knew" before writing captions.That led to a simple guiding question:What do I want the checkpoint to explain, and what do I want the trigger to own?Applying that idea consistently resulted in cleaner datasets and more modular LoRAs, which ultimately motivated a complete retraining of this model.<h2 id="about-version-2">About Version 2</h2>Version 2 is not deprecated or "incorrect."It reflects the captioning practices I was using at the time and remains an important reference point in the development of this methodology. Many of the ideas used in Version 3 only became apparent after comparing results across multiple generations of datasets.If you're interested in the evolution of the project, I encourage you to compare both versions—they represent two different approaches to the same training problem.<h2 id="mmanara_style_v2-european-comic-illustration-style-lora-(sdxl)">MManara_Style_v2 — European Comic Illustration Style LoRA (SDXL)</h2>Inspired by the emotional framing, compositional restraint, and romantic visual storytelling traditions associated with European comic illustration and the work of Milo Manara.This LoRA is not intended as a direct reproduction of any specific artist’s work, panels, or compositions. Instead, the project focused on capturing broader stylistic qualities:<ul><li>elegant contour economy</li><li>selective detail emphasis</li><li>restrained emotional expression</li><li>voyeuristic and observational framing</li><li>romanticized realism</li><li>cinematic stillness</li><li>watercolor softness</li><li>atmospheric environmental composition</li></ul>The training approach intentionally explored the idea of “absence as detail” — using implication, simplified rendering, and compositional suggestion rather than exhaustive detail.<hr /><h2 id="training-philosophy">Training Philosophy</h2>Rather than aggressively teaching:<ul><li>exact faces</li><li>repeated compositions</li><li>dense anatomical detail</li></ul>…the dataset and captioning workflow emphasized:<ul><li>emotional framing</li><li>posture and gaze</li><li>interpersonal spacing</li><li>compositional asymmetry</li><li>narrative atmosphere</li><li>selective rendering logic</li></ul>Captions were manually curated and standardized to reinforce:<ul><li>visual economy</li><li>contour discipline</li><li>soft watercolor shading</li><li>understated emotional tension</li><li>observational composition</li></ul>The goal was to encourage SDXL to learn not only surface aesthetics, but also aspects of cinematic emotional staging and illustrative restraint.<hr /><h2 id="training-details">Training Details</h2><ul><li>SDXL native training</li><li>~100 curated and manually reviewed images</li><li>manually corrected captions</li><li>trained with kohya_ss</li><li>base model: Juggernaut XL Ragnarok</li><li>rank: 24</li><li>alpha: 12</li><li>cosine scheduler</li><li>trained to ~3000 steps</li></ul>Additional caption reinforcement was used for:<ul><li>line stability</li><li>contour confidence</li><li>silhouette readability</li><li>controlled line hierarchy</li></ul><hr /><h2 id="recommended-prompting">Recommended Prompting</h2>Trigger:<pre><code></code></pre><pre><code>MMStyle</code></pre>Recommended reinforcement:<pre><code></code></pre><pre><code>European comic illustration</code></pre>Additional helpful style phrases:<pre><code></code></pre><pre><code>observational realism,
quiet sensuality,
voyeuristic framing,
restrained emotional expression,
cinematic stillness</code></pre>Suggested weight range:<pre><code></code></pre><pre><code>0.6 – 0.9</code></pre><hr /><h2 id="strengths">Strengths</h2>This LoRA tends to perform especially well with:<ul><li>romanticized environments</li><li>medium and environmental shots</li><li>atmospheric interiors</li><li>Mediterranean-style lighting</li><li>candid interpersonal scenes</li><li>cinematic compositions</li><li>elegant silhouette-driven posing</li></ul>It also transfers surprisingly well into:<ul><li>landscapes</li><li>sci-fi environments</li><li>market scenes</li><li>atmospheric architecture</li><li>emotionally suggestive compositions</li></ul><hr /><h2 id="notes">Notes</h2>This project is intended as:<ul><li>stylistic exploration</li><li>tribute</li><li>compositional study</li><li>emotional framing research</li></ul>The emphasis throughout training was on:<ul><li>atmosphere</li><li>selective rendering</li><li>emotional geometry</li><li>visual storytelling</li><li>implied form</li></ul>rather than direct visual duplication.

MManara_Style_v2

Updated captioning strategy to better allow the trigger to carry all of the style information. This version should be more flexible, cleaner triggering, and more consistently "Manara".

Files with SHA256: e06d15170616395acefa072d374b458c879a1eb379f2fc78608120fc57adffbc