Testing SD 3.5 Model in ComfyUI - is it Finally Better Than Flux?

Поділитися
Вставка
  • Опубліковано 29 лис 2024

КОМЕНТАРІ • 23

  • @ragemax8852
    @ragemax8852 Місяць тому

    I'm using both and 3.5 is coming out way better the more I play with it. The skin texture looks better too. Right now, 3.5 is looking way better for me.

  • @ZakariaNada
    @ZakariaNada Місяць тому +6

    Flux is still leading the industry by far. Come on stability AI, do better than this!
    Thanks for the video!

    • @Smokeywillz
      @Smokeywillz Місяць тому +2

      Um not anymore my friend. Flux is beautiful but a “dead end” model, it’s a distilled model meaning training a fine tune is not…practical. Sd3.5 is another flexible base model that shows great foundation if your not an ai casual. Flux doesn’t have much longer before the community makes sd3.5 absolutely overpowered. It’s undistilled and easy to train, that’s it’s selling point, not being “the best ever” model out the box

    • @nahlene1973
      @nahlene1973 Місяць тому +1

      Flux images are great out of box, but it is only good for generating images of certain needs, and it lacks the flexibility and expandability as it's distilled (not possible to train checkpoint models), a bit like a "MidJourney" that runs locally.

  • @lastreligiondfw
    @lastreligiondfw 12 днів тому

    what model of python, cuda and torch are you running?

    • @Aiconomist
      @Aiconomist  11 днів тому +1

      Cuda 12.4, Pytorch 2.5.1, Python 3.11.9

  • @shuntera
    @shuntera Місяць тому +3

    The Flux impressions of the women seem all to be of professional models with glamorous lips eyes and hair. SD3.5 seems to be better at generating more normal looking women. Though one of the prompts asked for blonde hair and they both looked like redheads to me.

  • @frankieb.fernandez79
    @frankieb.fernandez79 Місяць тому

    Any idea if we have to download the VAE and Unet to make this thing work?

    • @Aiconomist
      @Aiconomist  Місяць тому

      You only need the clips from SD3 medium

  • @DxBang3D
    @DxBang3D Місяць тому +2

    07:07 "photo of redhead.... she has blonde hair..."
    Are you trying to confuse the AI?

  • @bentontramell
    @bentontramell Місяць тому +4

    Flux chin has entered the room 😊

  • @danilsi6431
    @danilsi6431 Місяць тому +1

    The model is marvellous, beautiful, admirable, but it can't draw a person

  • @stefanvozd
    @stefanvozd Місяць тому +6

    "notable anatomical issues" 🤣

  • @krectus
    @krectus Місяць тому +5

    5:04 yeah the lighting is a bit off......oh and also she's missing an entire hand!

  • @Sujal-ow7cj
    @Sujal-ow7cj Місяць тому

    It's a base model, i think it will beat flux in time and i guess it's fast too

  • @MrArrmageddon
    @MrArrmageddon Місяць тому

    I feel like its worse than Flux without user models etc. First of all it does not support higher resolutions and I I think it performs slightly worse. It's not a bad model it's just SD 1.5 is better for being able to do uncensored or highly customized content. And Flux is all around better and can do words etc. Putting SD 3.5 around the SDXL middle ground of it's alright but it's more about what can it do later that matters. Honestly the real bread winner will be some form of future model that is extremely uncensored maybe even trained on nudity. Yes sex sells it pushes Industries. Reason why SD 1 in a lot of way can still be the best and has received the most love. Well I guess SD 1 is also very "easy" to train like light weight etc. But I hope 3.5 can get some amazing user models it may surpass FLUX if it can.

  • @alcocat
    @alcocat Місяць тому

    Answer is NO, in many ways

  • @amykpop1
    @amykpop1 Місяць тому +2

    they still can't manage to generate hands, man what are they even doing at this point? you're not even comparing it to flux's best and latest model and it's still failing. 🤣🤣🤣

    • @Smokeywillz
      @Smokeywillz Місяць тому +3

      It’s a BASE model that is UNDISTILLED with high levels of flexibility and flux can’t be trained. Put 2 and 2 together.

  • @geoffphillips5293
    @geoffphillips5293 Місяць тому

    Seems to be considerably worse than SDXL in terms of image quality.

  • @metasamsara
    @metasamsara Місяць тому +1

    imagine releasing a bad checkpoint just to control DRM.

    • @metasamsara
      @metasamsara Місяць тому

      the difference between flux training and sd3.5 training is that flux training data wasn't stolen from facebook profiles.

  • @ronbere
    @ronbere Місяць тому +1

    no...its horrible