<h1 id="pyradark">pyradark</h1><p>A character LoRA for generating a dark alternate version of Homura while still making use of the checkpoint’s existing character knowledge.</p><p>Rather than replacing the base model’s understanding entirely, <strong>pyradark</strong> is designed to reinforce the character while adding the dark alternate costume and related details. It can also work as a useful safety net when you are not fully confident that your checkpoint can reproduce regular Homura consistently on its own.</p><h2 id="activation">Activation</h2><p><strong>Trigger</strong><br /><code>pyradark</code></p><p><strong>Appearance</strong><br /><code>short hair, purple hair, tiara, red eyes, chest jewel, gem</code></p><p><strong>Outfit</strong><br /><code>alternate costume, dark persona, armors, elbow gloves, purple thighhighs, micro shorts, suspenders hanging</code></p><h2 id="recommended-strength">Recommended Strength</h2><p><strong>0.6–0.8</strong></p><p>This range is generally recommended for keeping the character and costume details stable without pushing the LoRA too aggressively.</p><h2 id="why-use-this-lora">Why Use This LoRA?</h2><p>The main purpose of pyradark is to let you generate the dark alternate costume while still taking advantage of the checkpoint’s existing knowledge of Homura.</p><p>The LoRA was trained at <strong>rank 32</strong>, which helps preserve smaller costume and character details with relatively high fidelity.</p><p>Since the silhouettes in the training images are mostly based on Homura’s standard visual structure, the LoRA can also help reinforce her identity when the checkpoint alone struggles to reproduce her reliably.</p><p>In that sense, it can be useful not only as an alternate-costume LoRA, but also as a kind of character consistency support.</p><h2 id="limitations-usage-notes">Limitations / Usage Notes</h2><p>Because this is a <strong>rank 32 LoRA</strong>, combining it with other LoRAs can cause instability or visual degradation more easily.</p><p>When using multiple LoRAs together, consider lowering the LoRA strength and reducing the amount of prompt conditioning. Some manual balancing may be necessary depending on the checkpoint and other LoRAs being used.</p><h2 id="training-information">Training Information</h2><p><strong>Base Model:</strong> WAI Illustrious v16<br /><strong>Epochs:</strong> 10<br /><strong>Training Steps:</strong> 3000<br /><strong>Optimizer:</strong> AdamW 8bit<br /><strong>Batch Size:</strong> 1</p>

Files with SHA256: 1f52c35d7f5b7d8eacb28d34dd049b48ef047873becf659df65436ae9842cba5