{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreieptcsllvwuwl2pl2zdverlha7hvr67fgvgkyppctpd7gubxats3m",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mnl4fh4q33i2"
  },
  "path": "/t/problems-with-duplication/176137?page=2#post_36",
  "publishedAt": "2026-06-05T20:36:52.000Z",
  "site": "https://discuss.huggingface.co",
  "textContent": "Hi John\n\nOkay, I’ll try it. I’ve been investigating, and the impainting models do something very curious. For example, when there’s a watermark so transparent that you can see all the action behind it, the models don’t recognize it as such. They don’t see that it’s actually visible, and they don’t need to erase anything; simply matching the tone of (in this case) a face or body with the video’s tone is enough, and they don’t need to erase anything.\n\nI understand that when something in the video is completely covered and the model can’t see it, it tries to imagine, but when a logo is so transparent that it doesn’t need to erase anything, IT DOES THAT TOO. It’s also strange that, until now, no one has improved or invented a model that recognizes that a watermark is transparent, sees what’s behind it, and simply tries to match the watermark’s tone with the colors of the original video.\n\nThese models are open source and can be improved, but I imagine that programming something like this requires a lot of knowledge, which is why no one has done it yet.\n\nI tested your entire model, and when there’s a face or something moving quickly in the background, it doesn’t work because it erases everything and tries to imagine it, which is why it creates something distorted. I researched it, and what needs to be done with faces after diffuraser is to capture a clean frame and use face swap on the entire video. THERE IS NO OTHER WAY AS OF TODAY, EVEN IF THE WATERMARK IS TRANSPARENT.\n\nEXAMPLE IMAGE: I’m sending you a screenshot. If you look at it, you’ll see that the watermark is very, very faint; you can see all the action in the background. I tested the model with all the settings, and it always erases all the action and tries to imagine it, and it imagines it incorrectly, all distorted, as I said before. THE MODEL DOESN’T SEE THE TRANSPARENCY; IT DOESN’T SEE THAT IT DOESN’T NEED TO ERASE ANYTHING, JUST MATCH THE COLOR TONES. IT’S A SHAME.",
  "title": "Problems with duplication"
}