<h1 id="heading-2578">🌟Helga Grey🌟</h1>📢Please read the version notes for SDXL v1.2! 📢📢(Helga Grey SDXL and updated SD15 LoRAs now available) 📢also known as Goldie Baby, is a Ukrainian-born adult film actress who gained popularity for her stunning looks and alluring performances. She made a name for herself in the industry with a few films released between 2015 and 2016, using various stage names such as Goldie, Goldie B., and Ms. White-Kitten. Known for her piercing blue eyes and natural 5'1" figure, Helga's beauty is complemented by her blonde hair and a 34-inch butt. Her real name is Karen G., and she was born on September 26, 1994, in Kiev, Ukraine. Despite her brief stint in the adult film industry, Helga remains a private person, and little is known about her personal life.<h1 id="heading-2579">💡Generation Advice💡</h1><h3 id="heading-2580">🔧Prompt and Sampling (Using auto1111 or similar UX) 🔧</h3><ul><li>Most images have been generated with LORA weighted between 0.65 to 0.75.</li><li>Lower weighting needed often for headshots.</li><li>Samples done with CyberRealistic v20 with sampling steps above 20, generally ~40 using Euler a, and DPM++ SDE Karras.</li><li>When testing a prompt, I often do a low lowres gen of multiple images, then choose the seed I like and upscale.</li></ul><h3 id="heading-2581">🚀Upscaling🚀</h3><ul><li>Portrait and close-ups work great using Hires. fix and 8X NMKD-Faces at 0.15 denoise strength.</li><li>Sometimes I just send it to Ultimate SD upscale after a quick render and run 2x at a scale of two, but often just the Hires. fix is good enough to capture enough details for the subject matter.</li><li><a target="_blank" rel="ugc" href="https://github.com/Coyote-A/ultimate-upscale-for-automatic1111">Ultimate SD upscale</a> script used @ ~.15 - .4 denoising using 8x NMKD Faces or 4X UltraSharp redraw upscalers.</li></ul><h1 id="heading-2582">🗒️Side Notes🗒️</h1><ul><li>In contrast to v1.0 and going forward, this LORA should be more ready to produce subject matter details during txt2img. Change your workflow accordingly.</li><li>The prompts continue to evolve. CyberRealistic is a greatly flexible model for being realism focused and still capturing abstract concepts. I still think they have trumped the SD 1.5 game. 👑</li><li>Feedback appreciated. Please leave a review if you've used this LORA. 👍</li><li>Leave a comment if you are having issues getting good output, I'll try to respond.</li></ul><h1 id="heading-2583">🏗Change Log🏗</h1><h2 id="heading-9">SDXL v1.2</h2><ul><li>different trigger to v1.0, use: ohwx woman</li><li>far more flexible model than XL v1.0</li><li>trained at 1216x1216px, top native resolutions are 1216x1216, 960x1408 and 1472x960</li><li>not recommended to go below 1216x1216, but if you do using resolutions smaller than that will require you to up the strength</li><li>long distance shots will also require you to up the strength</li><li>you know what, just start at 1.1 strength...</li></ul><h2 id="heading-57">SDXL v1.0</h2><ul><li>+1024px dataset only.</li><li>Still tweaking, will revisit when my process is more refined.</li></ul><h2 id="heading-57">SD15 v1.2</h2><ul><li>Completely reworked, from scratch +1024px dataset only. For future use with SD15 and SDXL.</li></ul><h2 id="heading-2584">1.1</h2><ul><li>Slightly expanded dataset, added far more detail to face.</li><li>Changed to new LORA single epoch creation methodology to get more consistent results and settings mostly usable across LORAs.</li><li>Only tested with Cyberrealistic v2.0, results may vary.</li></ul><h3 id="heading-2585">1.0</h3><ul><li>First gen, an incredibly ridiculous amount of steps. Can produce impressive results...sometimes, with way too much fiddling to get it. My first LORA attempt so not surprised.</li><li>Old release description: Ukrainian model for all you men (and women) of culture. This is my first LORA ever so please understand it is probably not ideal, but I think it is good enough to share. She is one of my favourites, so I'm sure I'll upgrade her in due time.</li></ul>

Helga Grey (Goldie Baby) + XL

dataset redone from scratchusing a +1024px baseattempted to reduce steps in training to prevent overcooking

Files with SHA256: 5e62429ae833025ea13644fc9bb858f6c51fc4efc5cb9479d3f415321885510c