{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreie6luhdyz7fzgrengxotd3bmgfo7l65iqxwhqivryrtaexces6coy",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mnohsporfgw2"
  },
  "path": "/t/problems-with-duplication/176137?page=3#post_49",
  "publishedAt": "2026-06-07T03:58:16.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "https://huggingface.co/spaces/John6666/ltx2.3-icedit-insight-rm"
  ],
  "textContent": "> –disable-white-mask\n\nOh… I missed that option… It doesn’t look like that option will magically solve everything, but for now I made one final version where it can be used just in case: https://huggingface.co/spaces/John6666/ltx2.3-icedit-insight-rm\n\n* * *\n\nLLM-generated explanation:\n\nThe `disable-white-mask` option turns off the pipeline’s default white conditioning mask. In this workflow, that mask seems to help preserve the original video structure while applying the IC-LoRA edit. Turning it off may give the model more freedom to modify stubborn areas, so it might help when a centered translucent watermark is barely changing at all.\n\nHowever, this is not a guaranteed watermark-removal switch. It does not specifically detect the watermark, and it does not create a clean ground-truth background. It only changes how strongly the edit is constrained by the original video conditioning. So the expected benefit is: the watermark may change more visibly than before. The expected limitation is: if the model cannot infer a clean replacement for that region, the watermark may still remain, or only become partially blurred.\n\nThe trade-off is quality and stability. With `disable-white-mask` enabled, the model can also modify things we want to preserve, such as players, grass texture, field lines, jerseys, text, or camera-motion consistency. For that reason I added it as a last-try preset rather than the default.\n\nAlso, I do not have your exact test video, so I cannot reproduce your specific case on my side. I can only test similar examples or synthetic/HF examples. Based on the tests so far, this model does work on some simple watermark examples, but it does not look very promising for a large centered translucent sports-broadcast watermark like this. If the new last-try preset still leaves the watermark almost unchanged, then this is probably a model/task limitation rather than just a missing UI option.\n\nIf possible, please test only a short 25-frame clip first. If you can share a short sample clip or the generated probe/debug zip, I can look at the exact behavior more accurately.",
  "title": "Problems with duplication"
}