<h1 id="ova07-v2-pocket-overall-skirt">OVA07 v2 – Pocket Overall Skirt</h1><h3 id="improved-garment-structure-cleaner-training-data-and-special-garment-poses">Improved Garment Structure, Cleaner Training Data &amp; Special Garment Poses</h3>OVA07 v2 is an improved version of my Pocket Overall Skirt / Cotton Overall Mini Skirt Concept LoRA.This version was rebuilt with a cleaner and more carefully structured training dataset, with particular attention to garment construction, suspender hardware, front bib structure, and controllable garment demonstration poses.The main goal of v2 is to improve both normal garment reproduction and the special poses involving the front bib and suspender straps.<hr /><h2 id="main-trigger">Main Trigger</h2><code>ova07_pocket_overall_skirt</code>Use this trigger for the basic garment concept.The LoRA is designed to reproduce a casual cotton overall mini skirt featuring:<ul><li>Front bib panel</li><li>Large front bib pocket</li><li>Two suspender straps</li><li>Metal suspender hooks and attachment hardware</li><li>A-line mini skirt silhouette</li><li>Casual cotton fabric construction</li><li>Multiple color variations, including navy, pastel pink, and black</li></ul><hr /><h1 id="what-was-improved-in-v2">What Was Improved in v2?</h1><h2 id="1.-cleaner-training-images">1. Cleaner Training Images</h2>The original training material contained some images originating from fashion product pages and other composite layouts.For v2, the dataset was carefully cleaned so that the LoRA focuses much more strongly on the actual garment itself.Where possible, the following elements were removed from the training images:<ul><li>White webpage margins</li><li>Product-page layouts</li><li>Social-media-like framing</li><li>Unnecessary text areas</li><li>Multi-image compositions</li><li>Collage-style layouts</li><li>Large blank spaces surrounding the photograph</li></ul>Composite source images were separated and cropped into individual training images whenever appropriate.This should reduce unwanted generations such as:<ul><li>Collages</li><li>Contact sheets</li><li>Social media post layouts</li><li>Website screenshots</li><li>Large white borders</li><li>Multi-panel images</li><li>Unwanted page-like compositions</li></ul><hr /><h2 id="2.-more-accurate-captions">2. More Accurate Captions</h2>One of the most important changes in v2 is the caption structure.Some images in the earlier dataset contained action descriptions that did not perfectly correspond to the visible garment state.For v2, captions were rewritten so that special pose triggers are only applied when the corresponding action is actually visible in the training image.Normal garment images are therefore trained primarily with:<code>ova07_pocket_overall_skirt</code>while special garment states use additional dedicated triggers.This is intended to reduce confusion between:normal garment appearance and special garment demonstration poses.<hr /><h1 id="special-garment-pose-triggers">Special Garment Pose Triggers</h1>OVA07 v2 introduces a more structured trigger system for special garment configurations.<h2 id="1.-lifted-folded-bib">1. Lifted / Folded Bib</h2>Trigger:<code>ova07_lifted_bib</code>Designed for configurations where the front bib panel is lifted, flipped, or folded away from its normal position.Example:<code>ova07_pocket_overall_skirt, ova07_lifted_bib, front view, lifted front bib, bib panel flipped upward and outward, asymmetric bib configuration, suspender hardware visible, photorealistic fashion photography</code>This trigger focuses on the unusual movement and folding behavior of the front bib rather than simply reproducing the garment in its normal worn state.<hr /><h2 id="2.-one-suspender-strap-undone">2. One Suspender Strap Undone</h2>Trigger:<code>ova07_one_strap_undone</code>Designed for asymmetric configurations where one suspender strap is loosened or detached while the opposite side remains connected.Example:<code>ova07_pocket_overall_skirt, ova07_one_strap_undone, front view, one suspender strap detached, opposite suspender strap remains attached, asymmetric suspender configuration, metal hook hardware visible, photorealistic</code>This can be useful for fashion styling, garment construction demonstrations, and more dynamic overall-skirt poses.<hr /><h2 id="3.-pulling-the-suspender-strap">3. Pulling the Suspender Strap</h2>Trigger:<code>ova07_pulling_strap</code>Designed for interactive poses where the model holds and pulls a loose suspender strap.Example:<code>ova07_pocket_overall_skirt, ova07_one_strap_undone, ova07_pulling_strap, front view, hand gripping detached suspender strap, pulling loose strap outward, opposite strap remains attached, metal suspender hardware visible, photorealistic fashion photography</code>This trigger is intended to improve the relationship between the model's hand, the suspender strap, and the garment hardware.<hr /><h1 id="recommended-basic-prompt">Recommended Basic Prompt</h1>For a normal garment generation:<code>ova07_pocket_overall_skirt, 1girl, front view, cotton pocket overall mini skirt, front bib pocket, two suspender straps, metal suspender hardware, casual fashion photography, photorealistic</code>For testing the special poses, I recommend starting with the dedicated trigger first and keeping the rest of the prompt relatively simple.For example:<code>ova07_pocket_overall_skirt, ova07_lifted_bib, 1girl, front view, standing, photorealistic fashion photography</code>Then gradually add additional pose descriptions only if necessary.<hr /><h1 id="recommended-lora-strength">Recommended LoRA Strength</h1>A good starting range is:0.7 – 0.9Recommended starting point:0.8The optimal value may vary depending on the checkpoint, prompt, sampler, and image composition.For special garment poses, increasing the LoRA strength does not always produce a stronger pose. If the garment structure becomes unstable, try returning to approximately 0.8 and adjusting the action trigger or prompt instead.<hr /><h1 id="suggested-generation-settings">Suggested Generation Settings</h1>A useful starting configuration:<ul><li>LoRA Strength: 0.8</li><li>Steps: 25–30</li><li>CFG Scale: 5.5–6.5</li><li>Sampler: DPM++ 2M</li><li>Scheduler: Karras</li><li>Portrait-oriented resolution recommended</li></ul>These are only starting recommendations and can be adjusted depending on your checkpoint.<hr /><h1 id="important-note-about-special-poses">Important Note About Special Poses</h1>The special garment poses are more difficult than normal clothing generation because they require the model to understand the relationship between:bib panel → suspender strap → metal hook → attachment point → fabric movementThe v2 dataset was specifically reorganized to improve these relationships.However, some original reference images naturally contain more than one garment action at the same time. For example, a lifted bib may also appear together with an undone suspender strap.Because of this, combining:<code>ova07_lifted_bib</code>and<code>ova07_one_strap_undone</code>may sometimes produce more reliable results than requesting the lifted-bib state completely independently.Similarly, the pulling-strap examples are closely related to the undone-strap configuration, so the following combination may work particularly well:<code>ova07_one_strap_undone, ova07_pulling_strap</code>Experimenting with individual and combined triggers is recommended.<hr /><h1 id="negative-prompt-suggestions">Negative Prompt Suggestions</h1>If unwanted webpage or collage-style compositions appear, try:<code>collage, contact sheet, image grid, multiple panels, multiple images, social media post, website screenshot, webpage, user interface, white border, white margin, text, logo, watermark</code>For normal single-image fashion photography, adding:<code>single model, single standalone photograph, one continuous scene, image fills the entire canvas</code>to the positive prompt may also help.<hr /><h2 id="v2-summary">v2 Summary</h2>Compared with the original version, OVA07 v2 focuses on:<ul><li>Cleaner single-photo training images</li><li>Reduced webpage / social-media layout contamination</li><li>Removal of unnecessary white margins and composite layouts</li><li>More accurate image-specific captions</li><li>Better separation between normal and special garment states</li><li>Dedicated triggers for special bib and suspender configurations</li><li>Improved front bib structure</li><li>Improved suspender strap relationships</li><li>Better representation of metal suspender hardware</li><li>More controllable garment demonstration poses</li></ul>The goal of this version is not only to reproduce the overall skirt, but also to make the garment behave more naturally when its bib and suspender straps are moved, loosened, lifted, or demonstrated in unusual configurations.Feedback, example generations, and prompt experiments are always welcome.

Files with SHA256: 41e58874fdc6e80ec1f1dbba8274f80ba89ffb50cca224da1e5f27e6c487de95