<blockquote>⚠️ WARNING: HIGH RESOURCE / EXTREME RANK (21 GB) This is NOT a lightweight adapter. This is a Full High-Rank Expansion (r=4096) trained via custom high-throughput architecture (Viking Engine). It expands the base model's attention capacity to learn complex multi-layered refractive glass, clear acrylic physics, internal mechanics, and anatomical X-ray optical depth.</blockquote>Key Features:<ol><li>True Layered Transparency: Renders real optical refractions, internal copper coils, circuit boards, and gears beneath glass/acrylic shells without melting into visual noise.</li><li>X-Ray &amp; Anatomical Physics: Seamlessly blends internal bone structure, cyan/neon luminescence, and clothing layers.</li><li>High Character ID Compatibility: Works exceptionally well in dual-LoRA setups! Preserves facial features and character identity even at lower weights while retaining full transparency physics on objects and apparel.</li></ol>Recommended Weights &amp; Settings:<ul><li>Standalone / Pure Cyberpunk Tech: <code>1.0</code></li><li>Merged with Character LoRAs / Commercial Product Shots: <code>0.60 – 0.75</code> (Optimal sweet spot: 0.70)</li><li>Sampler: <code>res_multistep</code> (or <code>euler_ancestral</code> / <code>dpmpp_2m</code>)</li><li>Scheduler: <code>simple</code></li><li>Steps: <code>50</code></li><li>CFG Scale: <code>4.0</code></li><li>Base Resolution: <code>2160 x 3840</code> (Upscaled / High-res) or native dataset resolution.</li></ul><ul><li>Latent Image Scale: <code>720 x 1280</code> (Up-scaled via Lanczos / Area to <code>2160 x 3840</code>)</li></ul>Here you will find tools for training models and preparing datasets<a target="_blank" rel="ugc" href="https://github.com/OrakulStudio">GitHub: https://github.com/OrakulStudio</a>

Files with SHA256: a94f2d00be3f7e463e5de44c9b5298651882e05b61b5f49494e77bbcafbaaf80