The latest articles on DEV Community by Галина Садовская (@__678da4d0). https://dev.to/__678da4d0 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3678686%2F9fddf1ce-670b-4af5-98d5-65623ca8382c.png https://dev.to/__678da4d0 en Галина Садовская Thu, 25 Dec 2025 19:31:23 +0000 https://dev.to/__678da4d0/ai-look-alike-search-for-of-creators-need-advice-on-better-face-models-58dc https://dev.to/__678da4d0/ai-look-alike-search-for-of-creators-need-advice-on-better-face-models-58dc <h1> AI Look-Alike Search for OF Creators — Need Advice on Better Face Models </h1> <p>I’m currently building an <strong>AI-based face similarity (look-alike) search for OF models</strong> as part of a real-world side project.</p> <p>The dataset contains <strong>100,000+ public OF model images</strong>, and the goal is to help users discover <strong>visually similar OF models</strong> based on facial features rather than usernames or text-based search.</p> <p>This is <strong>not identity verification</strong> — the focus is purely on <strong>visual similarity</strong>.</p> <h2> What I’m Building (Quick Overview) </h2> <ul> <li>Users upload an image (reference photo / celebrity image)</li> <li>The system finds <strong>OF models with similar facial characteristics</strong> </li> <li>Results are ranked using <strong>face embeddings + vector similarity search</strong> </li> <li>Everything currently runs on <strong>CPU</strong>, but I’m considering a move to <strong>GPU</strong> for scale and experimentation</li> </ul> <h2> What I’m Building (More Detail) </h2> <p>The system allows users to upload an image and receive a list of <strong>OF models with similar facial characteristics</strong>.</p> <p>The intent is to support <strong>visual discovery</strong>, where perceived similarity matters more than exact identity matching.</p> <h2> Key Constraints </h2> <ul> <li><p><strong>Similarity over identity</strong><br><br> The system ranks faces by perceived similarity (look-alike matching), not by strict identity verification.</p></li> <li><p><strong>Low tolerance for false positives</strong><br><br> Returning visually different faces as “similar” is more harmful than missing a potential match.</p></li> <li><p><strong>Real-world images</strong><br><br> The dataset consists of non-studio images with varying lighting, poses, resolutions, and overall quality.</p></li> <li><p><strong>Scalability</strong><br><br> The solution needs to scale beyond <strong>100k+ images</strong> without significant drops in accuracy or performance.</p></li> </ul> <h2> Current Pipeline (CPU-Based) </h2> <p>At the moment, the entire pipeline runs on <strong>CPU only</strong>.</p> <p>The setup looks like this:</p> <ul> <li>Face detection and alignment</li> <li>Feature extraction using a pre-trained face model</li> <li>Storing embeddings in a vector index</li> <li>Nearest-neighbor search using cosine similarity</li> </ul> <p>At this scale, the system works reasonably well, but both <strong>accuracy and performance</strong> are starting to become limiting factors.</p> <h2> Current Model Setup (InsightFace) </h2> <p>Face embeddings are currently generated using <strong>InsightFace</strong>, specifically the <strong><code>buffalo_l</code></strong> model bundle.</p> <p>The pipeline includes:</p> <ul> <li>Face detection and alignment via InsightFace</li> <li>Feature extraction using the <code>buffalo_l</code> model</li> <li>Embeddings stored for similarity search</li> <li>Cosine similarity for ranking similar faces</li> </ul> <p>This provides a solid baseline, but for <strong>look-alike matching</strong>, small inaccuracies are very noticeable.</p> <h2> Where the System Struggles </h2> <p>As the dataset grows, several issues become more apparent:</p> <ul> <li>Visually similar faces sometimes rank lower than expected</li> <li>Different individuals with shared facial traits can appear as false positives</li> <li>Lighting, pose, and image quality introduce noise</li> <li>CPU inference becomes a bottleneck during re-indexing and experimentation</li> </ul> <p>Because this is a <strong>look-alike use case</strong>, even small errors can significantly affect perceived quality.</p> <h2> CPU vs GPU — Is the Move Worth It? </h2> <p>I’m planning to migrate the pipeline to <strong>GPU-based inference</strong>, but I want to make sure the model choice justifies the move.</p> <p>Some of the questions I’m evaluating:</p> <ul> <li>Which face models provide the best results for <strong>visual similarity</strong>, not identity recognition?</li> <li>Does GPU inference unlock <strong>meaningfully better accuracy</strong>, or is it mainly a speed improvement?</li> <li>Are there models that are simply not practical to run on CPU at this scale?</li> </ul> <p>If I’m going to reprocess <strong>100k+ OF model images</strong>, I want to do it with the right model.</p> <h2> What I’m Looking for in a Better Face Model </h2> <p>I’m particularly interested in models that:</p> <ul> <li>Produce <strong>high-quality embeddings</strong> for similarity search</li> <li>Perform well on non-ideal, real-world images</li> <li>Scale efficiently beyond 100k images</li> <li>Benefit from GPU acceleration</li> <li>Can be fine-tuned (or perform well out of the box) for look-alike matching</li> </ul> <p>I’m open to both <strong>open-source</strong> and <strong>commercial</strong> solutions.</p> <h2> Real-World Context </h2> <p>This work is part of a discovery platform where users can upload an image and find visually similar <strong>OF models</strong> using AI-based face similarity.</p> <p>The project is called <strong>Explore.Fans</strong>, and face similarity search is one of its core components.</p> <p>👉 <a href="https://explore.fans" rel="noopener noreferrer">https://explore.fans</a><br><br> (Shared only for technical context.)</p> <h2> Questions for the Community </h2> <p>If you’ve worked with <strong>face similarity or face recognition models</strong> at scale, I’d really appreciate your input:</p> <ul> <li>Which models gave you the best results for look-alike similarity?</li> <li>Did GPU inference improve accuracy, or mostly performance?</li> <li>Any experience fine-tuning models for similarity-based ranking?</li> <li>Anything you’d avoid based on real-world experience?</li> </ul> <p>Thanks in advance — happy to share more details if helpful.<br> Have a wonderfull holiday!</p> <h2> References </h2> <ul> <li>InsightFace (DeepInsight): <a href="https://github.com/deepinsight/insightface" rel="noopener noreferrer">https://github.com/deepinsight/insightface</a> </li> </ul> ai deeplearning help