4 easy ways to train LoRAs for consistent characters in Leonardo, Runway, EverArt, and FullJourney

Поділитися
Вставка
  • Опубліковано 26 гру 2024

КОМЕНТАРІ •

  • @spacekitt.n
    @spacekitt.n 11 місяців тому +6

    for leonardo, generating with finetuned models works best with alchemy on, but prompt magic off. you can even give it a picture in image guidance with those settings (again make sure alchemy is on but prompt magic off) and it will more or less paste your face into the picture. also train with 2.1 (even though it suggests 1.5). 2.1 gives better results in my experience. ticking the high resolution toggle works fine too, even though it warns you it may give strange results (it never has for me)

    • @indiemusicvideoblog
      @indiemusicvideoblog 11 місяців тому

      So you don’t have to train it, you can just add 1 image in guidance?

    • @spacekitt.n
      @spacekitt.n 11 місяців тому +2

      @@indiemusicvideoblog no you still have to train it. it has to know the face to put in the output. like this vid says, it needs a good variety of angles and lighting...it will still do its best, but the more you can give it the better chances of success you'll have. and if your dataset has pictures very similar to the image in image to image, you'll probably have even more success. I dont know, others might have a different experience thats just what ive found, ive tested like 20 different models with every setting.

    • @indiemusicvideoblog
      @indiemusicvideoblog 11 місяців тому

      Fantastic. Thank you

    • @aivideoschool
      @aivideoschool  11 місяців тому

      This is super helpful, I'm going to give it a try later. Thank you!

    • @spacekitt.n
      @spacekitt.n 11 місяців тому +1

      @@aivideoschool yeah prompt magic adds another step in the process that completely changes things--im not sure how but whatever its doing, it mixes in your trained dataset with another dataset it seems. you can turn it on and get some good creative results but it just takes more tries to get something good that resembles things in your dataset. and if you do turn it on, keep the value low, like .3. either way, on or off it will take a bunch of tries to get something good--though when you do, its magical

  • @piggywahwah
    @piggywahwah 7 місяців тому

    wow fantastic, really love these in-depth walkthroughs, very nice vid :)

    • @piggywahwah
      @piggywahwah 7 місяців тому

      oh and that everart pricing is just off the scale??? wot the hell? It just says $50 starter now???

    • @aivideoschool
      @aivideoschool  7 місяців тому

      They went crazy with the pricing! I wonder if people are using it for AI influencers or something. That's too much for what it does for most users.

    • @piggywahwah
      @piggywahwah 7 місяців тому

      @@aivideoschool Yea, gutted,
      i even checked leonardo and it doesn't do img to vid? i click on the gen vid button and it has no option for load image??
      That looked really interesting.. doubly whammy 😣😣

  • @bunnystrasse
    @bunnystrasse 11 місяців тому +2

    So which one has the best quality of LORA?

    • @aivideoschool
      @aivideoschool  11 місяців тому +2

      In terms of best quality of likeness, I'd say EverArt has a slight edge over FullJourney, but FullJourney lets you experiment more. In terms of resolution quality, EverArt lets you upscale 2x or 4x but you only have a few upscales in the lowest tier plan I used.