<p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/6a17edba-8656-4315-88f2-32282cae373d/original=true/6a17edba-8656-4315-88f2-32282cae373d.jpeg" /></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!-portraits"><strong>SoReal! - Portraits</strong></h1><p>|| <a target="_blank" rel="ugc" href="https://civarchive.com/models/2357375/soreal-pov?modelVersionId=2657257">SoReal! - POV</a> || <a rel="ugc" href="https://civarchive.com/models/2368299/soreal-natural-bodies">SoReal! - Natural Bodies</a> || <a rel="ugc" href="https://civarchive.com/models/2384267/vacation-photos">Vacation Photos </a>|| </p><p></p><h2 id="overview"><strong>Overview</strong></h2><p>Capture the range of human beauty! This model is the next model in 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"><strong>Compatibility &amp; Usage</strong></h2><p>Due to it's small rank and layer filter, the model should have a minimal influence on the base models actual weights, 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! 'Portrait' or 'Portrait photography' appears very frequently in the captions so this may help you get closer to a more portrait-style image if your prompt is not very portrait-esque.</p><p>When using Z-Image Turbo, I use a strength of 1.</p><p></p><h2 id="limitations"><strong>Limitations</strong></h2><p>This isn't explicitly an NSFW model - it was trained on both male &amp; female anatomy, but there was nothing special done to it to improve adherence to NSFW concepts.</p><h2 id="future"><strong>Future</strong></h2><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"><strong>Dataset &amp; Training</strong></h2><p>Dataset of 3000 sourced from a variety of sources. Deduplication and Quality Scoring (through MANIQA) lowered the dataset to around 1200.</p><p>This model was trained on a dataset of 1200 images at a batch size minimum of 16.</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"><strong>Licensing</strong></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>

SoReal! - Portraits

<p>Improved fidelity and prompt adherence.</p>

Files with SHA256: dc497c86dcb4f84e90ef1e286479cf4793eb920122a938620269890cb8619c2e