<h1 id="snofrid-is-a-winter-muse-in-her-mid-20's.-ivjilfwyq">Snöfrid is a Winter Muse in her mid 20's.</h1>Snöfrid is not based on a real person. This Textual Inversion (TI) was designed to embody a timeless beauty inspired by cold elegance and resilience. Her features include high cheekbones, a softly retroussé nose, luminous pale skin, and a cold flush on her cheeks and lips, reminiscent of brisk winter air. Her striking icy white hair, curled with natural darker roots, adds depth to her character, while her almond-shaped eyes and soft cupid’s bow lips give her an air of ethereal serenity. Snöfrid’s understated elegance and natural vitality make her versatile for a wide range of scenarios—whether set in modern, fantasy, or futuristic environments.<h2 id="snofrid-ti-changelog:-mtz2bg2qu">Snöfrid TI Changelog:</h2><h3 id="v3-reduce-tokens-more-body-shape-details.-flym4enxi">v3 - Reduce tokens, more body shape details.</h3>After RTFM, I was able to use the <a target="_blank" rel="ugc" href="https://github.com/klimaleksus/stable-diffusion-webui-embedding-merge">embedding merger tool</a> to reduce the token use from 75 to 55. I also added details about body shape. This is still twice as many as I would prefer but c'est la vie. In XYZ testing, reducing below 55 made it prone to errors.<h3 id="v2-more-facial-feature-consistency.-upjm6vn2f">v2 - More facial feature consistency.</h3>Foundation: Version 2 was developed using outputs from v1, focusing on consistency and goal alignment. Generated faces were meticulously cherry-picked for quality and fidelity to the intended aesthetic.Image Preparation:<ul><li>Selected images were merged using <a target="_blank" rel="ugc" href="https://github.com/Gourieff/sd-webui-reactor">ReActor </a>and processed with Hi-Res Fix applied before downscaling back to standard training size.</li></ul>Training Process:<ul><li>A midpoint Textual Inversion (TI) was trained on SD <a target="_blank" rel="ugc" href="https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5">1.5 base</a>:</li><li>Steps 1–60: Base SD 1.5 model.</li><li>Steps 61–300: A balanced mix of my two models (Realistic, Illustrated) to refine general body shape and facial traits, blending realistic and anime 2.5 styles to emphasize diverse facial feature representation.</li></ul>Merging and Refinement:<ul><li>Final TI was merged with v1 at 0.8 weight.</li><li>A blend with the midpoint TI was applied at 0.9 weight, creating a cohesive representation of Snöfrid's facial and bodily features.</li><li>Finalization: The combined output, consisting of the original and midpoint TI, is released as v2, achieving greater versatility and feature coherence.</li></ul><h3 id="v1-initial-experimental-version.-qbyil4wkm">v1 - Initial experimental version.</h3>Initial experimental version that uses about 4-15x as many tokens compared to other face TI's that are trained on images. I'll aim in future versions to reduce this, but for now "it works", so I'm going to post it for completionist's sake. A lot of photography/realism models struggle with white hair and will often turn them blonde or even black. This model was not trained on actual images and will have fairly random faces aside from the key details mentioned above in bold. If you find a particular face you like, stick to that seed and adjust the prompt. I created this TI as a way to be referenced using wildcards of my <a target="_blank" rel="ugc" href="https://civarchive.com/collections/5588101">favorite faces</a>. Snöfrid was developed using the <a target="_blank" rel="ugc" href="https://github.com/klimaleksus/stable-diffusion-webui-embedding-merge">embedding merger tool</a> to focus exclusively on facial features, leaving backgrounds, body shapes, and art styles entirely up to the user's creative direction. By avoiding image-based training entirely, this Textual Inversion (TI) intentionally allows for some diversity in the generated faces, making it a flexible starting point for a variety of uses. This approach was chosen to sidestep common pitfalls seen in other embeddings, such as over-fitting to training artifacts, feature entanglement, poor dataset curation, and prompt hijacking—where a TI imposes its learned context too strongly, overriding user-specified prompts. The result is a versatile tool that maintains the core essence of Snöfrid's facial features while enabling dynamic adaptability across different styles, contexts, and artistic visions.

Files with SHA256: 3eaedb7c6aa69a8984da80979ff9ab74758df8f0d9adfa7e403a38e5a213f76c