• istanbullu@lemmy.ml
    link
    fedilink
    English
    arrow-up
    3
    ·
    2 months ago

    Is a reason to use SD3 when Flux is much much better, with a much better license?

    • Even_Adder@lemmy.dbzer0.comOP
      link
      fedilink
      English
      arrow-up
      2
      arrow-down
      1
      ·
      edit-2
      2 months ago

      I don’t think so. They’re going to have to do a lot better than a tutorial to win people back. That said, the two Flux models being distilled making them close to impossible to fine-tune sucks too.

      • clb92@feddit.dk
        link
        fedilink
        English
        arrow-up
        1
        ·
        edit-2
        2 months ago

        People have been training great Flux LoRAs for a while now, haven’t they? Is a LoRA not a finetune, or have I misunderstood something?

        • Even_Adder@lemmy.dbzer0.comOP
          link
          fedilink
          English
          arrow-up
          1
          arrow-down
          1
          ·
          2 months ago

          Last I heard, LoRAs cause catastrophic forgetting in the model, and full fine-tuning doesn’t really work.

          • clb92@feddit.dk
            link
            fedilink
            English
            arrow-up
            2
            ·
            2 months ago

            Oh well, in practice I’ll just continue to enjoy this (possibly forgetful and not-fully-finetunable) model then, that still gives me amazing results 😊

          • erenkoylu@lemmy.ml
            link
            fedilink
            English
            arrow-up
            1
            arrow-down
            1
            ·
            edit-2
            1 month ago

            quite the opposite. Lora’s are very effective against catastrophic forgetting, and full finetuning is very dangerous (but also much more powerful).

        • Even_Adder@lemmy.dbzer0.comOP
          link
          fedilink
          English
          arrow-up
          1
          arrow-down
          1
          ·
          2 months ago

          Those might just be LoRA merged models, not full fine-tuning. From what I heard, fine-tuning doesn’t work because the models are distilled. You’d have to find a way to undistill them to train them.