<h3 id="larp-(low-rank-adaptation-for-realistic-photography)">LARP (Low-rank Adaptation for Realistic Photography)</h3>This was a Krea 2 Hillclimb for ultimate realism and detail<ul><li>Trained to work exceptionally well when combined with <a target="_blank" rel="ugc" href="http://twotensors.ai">twotensors.ai</a>'s forensic cleaner when using for social media</li><li>Extremely high rank of 384 for maximum detail extraction (can also backfire for ~10% of prompts)</li><li>16k steps at BS=2 on ~6000 images captioned with GPT 5.6 Terra + other trainer modifications</li><li>Trained on RAW, generated on Turbo</li></ul>Generation Tips:<ul><li>This model is NOT for the GPU poor. Generations look significantly better the higher resolution you generate at (~3k optimal, it falls apart after that)</li><li>Stay at 8 steps, going over 8 steps can add TOO MUCH detail.</li><li>Prompts must be EXTREMELY detailed, most of the prompts in the post are from <a target="_blank" rel="ugc" href="https://www.promptlibrary.space/">https://www.promptlibrary.space/</a></li><li>Untested on Image editing workflows</li></ul>

Files with SHA256: 9b8fb4a5ef849a1e3665c3c04483c64c3d53996e87a6b647b1d60a1dd50a77a1