{
  "id": 19619433,
  "title": "Keeping identity intact in AI clothing swaps: what we learned building Outfia",
  "url": "https://urgent.news/2026/11/12/keeping-identity-intact-in-ai-clothing-swaps-what-we-learned-building",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-12T07:51:47.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/outfia/keeping-identity-intact-in-ai-clothing-swaps-what-we-learned-building-outfia-4fco"
  },
  "original_language": "en",
  "account": "Outfia is an AI-powered clothing swap tool that focuses on preserving the person's identity in the generated images. The model is designed to redraw only the clothing while keeping the face, pose, hands, lighting, and background unchanged. This constraint is crucial because a normal image-to-image prompt can allow the model to change too many aspects of the input photo, leading to a distorted result that fails the actual purpose of the clothing swap.\n\nTo achieve this goal, Outfia uses a \"preserve list\" that includes the face, hairline, skin tone, expression, body shape, pose, hands, background, lighting, and camera angle. The original image's crop and framing are also preserved. By focusing on only the clothing as the variable, the model can generate coherent results even when multiple garments are involved.\n\nThe garment parsing process is more critical than the prompt wording. Different clothing items (tops, bottoms, dresses, jackets, and full outfits) require different editing approaches. Tagging each garment item accordingly helps the model understand how to modify the image. Additionally, allowing up to three garment items in a single run keeps the output more coherent by avoiding chained generations that increase the risk of identity drift.\n\nOutfia currently uses two model families: Google Nano Banana 2 for Standard runs and OpenAI GPT Image for Pro runs. Each model has its strengths and weaknesses, and the choice depends on the input photo and the desired outcome. Standard is suitable for quick previews, while HD or Pro is recommended for final product shots.\n\nEvaluation of Outfia's results is manual and visual, rather than relying on text metrics. A side-by-side comparison of the original and edited images, along with checking specific details like face, hands, hairline, background, and garment features, ensures the best possible result. The before/after slider in the product allows users to see the failure modes and understand why a run may have failed.\n\nPrivacy is also a concern, as clothing photos are often personal. Outfia deletes uploads and results after 24 hours, does not use them for model training, and offers free results with a small watermark and limited resolution, while paid runs provide full-resolution images without watermarks. Failed runs are refunded automatically on all plans and credit packs.\n\nIn summary, Outfia is a valuable tool for those seeking a visual preview of how a garment might look on them without altering their overall identity. The key to its success lies in preserving the person in the edited image, with a focus on garment parsing, model choice, and manual evaluation.",
  "summary": "Most AI photo editors are optimized for \"make this look different.\" A clothing swap has the opposite requirement: the person has to look the same. If the face, hairline, body shape, hands, or background drift, the result stops being useful. It looks like a different person wearing the clothes. We built Outfia around that constraint. You upload a photo of yourself and up to three garment photos,…",
  "key_points": [
    "Outfia preserves identity by keeping face, pose, hands, lighting, and background unchanged.",
    "Manual visual evaluation ensures best results and identifies failure modes."
  ],
  "editors_take": "Prioritizing identity preservation through strict garment parsing and variable constraints transforms AI image editing from a generic generative process into a specialized tool for realistic virtual try-ons.",
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}