<h2 id="overview">Overview</h2>This training job focused on creating a LoRA extension to capture the visual and stylistic characteristics of a fictional AI character (loosely based on me) using the Z-Image De-Turbo diffusion architecture.<h3 id="model-and-architecture">Model &amp; Architecture</h3><ul><li>Base Model: <code>ostris/flux.2_klein_base_9b</code></li><li>Quantization: qfloat8 (model and text encoder)</li><li>Precision: bf16</li></ul><h3 id="training-configuration">Training Configuration</h3><ul><li>Training Steps: 3000</li><li>Optimizer: AdamW 8-bit</li></ul>The training emphasized a balanced content-to-style objective, suitable for flexible prompt usage.<h3 id="trigger-words">Trigger Words</h3><ul><li>Trigger word nekotaylor, woman, 1girl</li></ul>This model is designed to generate fictional, stylized human-like characters, not real person. The source was my face image, altered by AI (M2F). It focuses on imaginative art styles rather than photorealism. The faces, bodies, and features it produces are entirely artificial and do not depict or resemble any real individual, living or deceased.Ideal for artists and creators who want a character renders in a fantasy or conceptual art setting. Works well for portrait (not so good with full-body).

Files with SHA256: 570ff687051788e01cddbf78eec9883b4403f47ea388292347ca6765c1b1f64f