• Mixel@feddit.de
    link
    fedilink
    arrow-up
    8
    ·
    4 months ago

    To train a diffusion model that only outputs one image with difference is I think not possible you could do an image to image and then fix the seed so you would get a consistent result and then picking the nearest result that is nearly an identical copy

    • Hjalmar@feddit.nu
      link
      fedilink
      arrow-up
      9
      ·
      4 months ago

      It’s very much possible and indeed such a problem that it may be done by mistake if a large enough data set isn’t used (see overfitting). A model trained to output just this one image will learn to do so and over time it should learn to do it with 100% accuracy. The model would simply learn to ignore whatever arbitrary inputs you’ve given it