<p></p><h2 id="follow-me-on-patreon!"><a target="_blank" rel="ugc" href="https://www.patreon.com/Bitcrushed_Heart"><strong>Follow me on Patreon!</strong></a></h2><h1 id="soreal!-pov">SoReal! - POV</h1><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/2360068">[SoReal! Portraits]</a></p><h2 id="overview">Overview</h2><p>Reach your hands out for the stars! This model is the first of a series in Z-Image LORAs - aimed to bring diversity in both concepts and humanity itself to Z-Image.</p><p></p><h2 id="compatibility-and-usage">Compatibility &amp; Usage</h2><p>Due to it's small size and rank, the model should have a minimal influence on the base model, further improving compatibility with other LORAs across Base/Turbo and indeed other checkpoints.</p><p>'Trigger words' aren't real - don't ask for one, just prompt normally. If you want a hand (literally), use 'a man's hand' or 'a woman's hand', which should normally get you what you want.</p><p>I'll upload a full concept list soon to show the range of concepts the model has been trained on - not confirming that the model is able to reproduce them, though.</p><p>When using Z-Image Turbo, strengths between 0.95 and 1.5 work best in my experience for V1, and 0.9 - 1.2 for V2.</p><p></p><h2 id="limitations">Limitations</h2><p>Anatomy is still rough - planning one further additional training run to try and address this for NSFW concepts but may mean a split between generalisation model (v2) and a NSFW-special model (V2-NSFW).</p><p></p><h2 id="future">Future</h2><p>Future iterations of this model will see stronger prompt adherence, anatomy adherence and general composition and quality through +/- reinforcement learning.</p><p><strong>I am planning on finetuning Z-Image considerably with a model called 'SoReal!' (Or, alternatively, ZoReal!). However, I want it to be the best possible amateur finetune possible, to achieve this, I have:</strong></p><ul><li><p>1. Trained a custom quality model.</p></li><li><p>2. Trained a custom one-shot demographic model (height, weight, skin tone, ethnicity, age in years, body shape) with an average accuracy of 89% for top-confidence prediction using ConvNext-XL.</p></li><li><p>3. Finetuned wd-tagger-large-v3 on a large sample dataset of 50k hand-tagged images with human-assisted active learning.</p></li><li><p>4. Fed those tagged images (with quality, demographics and general labels) with the image metadata (incl. EXIF &amp; Camera Metadata) to Gemini 3 Flash for generating captions.</p></li><li><p>No over-trained LORAs baked in, no dramatic loss of generalisation, just a good, all-round, NSFW-ready, finetuned model.</p></li></ul><p><strong>I am now severely limited, however, by my compute and financial situation, so if you'd like to help make SoReal!, well, so real, </strong><a target="_blank" rel="ugc" href="https://www.patreon.com/Bitcrushed_Heart"><strong>then you can follow me on Patreon!</strong></a></p><p></p><h2 id="dataset-and-training">Dataset &amp; Training</h2><p>Dataset of 2500 sourced from a variety of sources. Deduplication and Quality Scoring (through MANIQA) lowered the dataset to around 1400. This model was trained on a dataset of 1500 images at a batch size minimum of 10. This means</p><p>This model was trained on a dataset of 1500 images at a batch size minimum of 10. Masked loss was implemented after roughly 40,000 samples (not steps) to improve anatomy &amp; concept adherence.</p><p>Validation loss was used with 10% of the dataset size to prevent overfitting while still maintaining strong concept adherence and generalisation.</p><p>Model was trained with AdamW through the Python adv-optm package.</p><p></p><h2 id="licensing">Licensing</h2><p>If you'd like to release a merge of this model, please contact me.</p><p></p><p>Made with &lt;3 By BitcrushedHeart</p>

Files with SHA256: 2954608b1954fe7fbf3885e14af268de5f56002fa5a424656ae237c6cca3929e