{
  "id": 20018311,
  "title": "Keeping a product's label real in AI video: a tracker with no green screen",
  "url": "https://urgent.news/2026/11/14/keeping-a-products-label-real-in-ai-video-a-tracker-with-no-green",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-14T17:40:32.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/zaqero_ai/keeping-a-products-label-real-in-ai-video-a-tracker-with-no-green-screen-5edp"
  },
  "original_language": "en",
  "account": "When creating ads for small brands using AI video models, the primary issue isn't the motion or lighting but the label. An incorrect label can mar the product's appearance. Zaqero AI, a company that creates short product videos for small businesses, addresses this problem. They build a tracker that maintains the label's authenticity in AI-generated videos.\n\n1. The process begins by not allowing any photo editing software to alter the product's image before it reaches the video model. An earlier version of their pipeline cleaned up the product photo with an image model, setting the product on a plain grey ground. This ensures the label the video model sees is already the image model's copy of it. The product is cut out of its background and placed on light grey with a margin, retaining every product pixel. The background is not re-created or restyled on the way.\n\n2. After a take comes back, a vision model is used to find the label's corners in the generated shot. It answers in JSON format, providing the position of the four corners of the label (and, for round objects like a jar, the left and right edges at the label's height). If the label is not clearly visible, the answer is \"not found,\" and the shot remains unchanged.\n\n3. Following the label without a green patch is the next step. Instead of using color keying, which failed on the label's own lettering, they follow the picture itself. Their patch tracker takes the label region from one reference frame and, on every other frame, searches for the movement and scale (and, for objects that turn, the rotation) that best match it, by normalised correlation. Two key points are that every frame is compared with the reference, never with the previous frame, and that the previous frame only provides a starting point for the search. This method prevents errors from accumulating.\n\n4. Instead of pasting a new label image over the old one, they flatten the label region out of the picture, fit a smooth curve through it to get the surface's own light, and treat what stands apart from that curve as print. Only the lettering is filled in from the same frame's clean surface and with its own fine noise. The label's edges, its light, and its grain remain in the footage. The new print, if requested by the brand, is then placed in the same location, at the same size and shape. They measure the shot instead of visually judging it to avoid a pasted appearance. They measure 'softness' and 'flicker' and adjust the print's appearance accordingly.\n\n5. They discovered a bug while measuring the shake, where every video had a stutter they had stopped noticing. The video model returns 24 frames per second, but they render at 30, causing every fifth frame to be a repeat. To prevent this issue, they now render at 24 frames per second.\n\n6. The label path is built and tested but not yet switched on in production. They have measured this process on one test shot with a stubbed vision step on their own rendering bridge. They will update the public once they have run it on real videos with actual numbers. The flattening is done bilinearly between four corners, which works well for labels seen nearly square-on, but reflections on glass, labels that leave or fold into the frame, and other complexities are not yet handled. They plan to write more once they have successfully implemented this system on real videos.",
  "summary": "If you make ads for a small brand with an AI video model, the first complaint is rarely the motion or the light. It is the label. A jar comes back with letters that are almost right, a logo that has melted a little, a line of small print that is now a different line of small print. To the brand that is not their product any more. We build Zaqero AI, which makes short product videos for small…",
  "key_points": [
    "Zaqero AI builds a tracker to maintain label authenticity in AI-generated product videos.",
    "Vision model finds label corners in generated shots, providing JSON format coordinates.",
    "Patch tracker matches label region across frames, preventing errors from accumulating."
  ],
  "editors_take": null,
  "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."
}