SD 3.5 Large Model: What’s the Best Sampler and Settings?

Поділитися
Вставка
  • Опубліковано 27 лис 2024

КОМЕНТАРІ • 21

  • @stereotyp9991
    @stereotyp9991 15 днів тому +1

    Can we please talk about the fact that all the the Roman numbers on the clock are correct???!!!! That is nuts!!!!!

  • @pawelthe1606
    @pawelthe1606 Місяць тому +2

    For me best settings are: sampler-euler, scheduler-sgm_uniform, steps-40, cfg-5.5. Great for realism and art.

    • @CodeCraftersCorner
      @CodeCraftersCorner  27 днів тому

      Thanks for sharing. I do the same. for steps, I do 20 and check the composition and then 40 for the final output. CFG: from 4.5 - 5.5 depending on the complexity of the prompt.

    • @admiralevan
      @admiralevan 19 днів тому

      thank you my friend

  •  Місяць тому

    Again very useful, thank you! I was scratching my head as to which parameter would do what... What is your personal opinion? What did you deduce from the test, what parameters do you think gave the best picture?

    • @CodeCraftersCorner
      @CodeCraftersCorner  27 днів тому +1

      Thank you! For me personally, i am using these: sampler: euler, scheduler: sgm_uniform, steps: 20 for (finding the composition), 40 (final image), cfg: 4.5 - 5.5 to make the model follow the prompt.

  • @francaleu7777
    @francaleu7777 Місяць тому

    thank you!!

  • @admiralevan
    @admiralevan 19 днів тому

    I'm getting a lot of all black images, do you know why? 40 steps, CFG scale 5.4, Euler, SGM Uniform, CLIP + T5, Clip l, Clip G, and T5XXL loaded, 1440x1440. Rendered in SwarmUI(uses comfyui background)

    • @CodeCraftersCorner
      @CodeCraftersCorner  16 днів тому

      Can you try at a lower resolution? Also, try to update ComfyUI to the latest version.

  • @devnull_
    @devnull_ Місяць тому

    TBH, I stick with Euler and whatever scheduler seems to work. Unless one has time to generate like 20 images with each possible combination of samplers and schedulers, it is pretty hard to know how it affect the prompt you are trying to craft. Some combinations clearly work better for some concepts, but I wish there was less choices of samplers and schedulers (in some sense, of course options are nice), and better prompt understanding instead.

    • @CodeCraftersCorner
      @CodeCraftersCorner  Місяць тому +1

      I, too, share the same opinion. It took me nearly 6 hours (with breaks in between) to generate all these images, and it’s just not practical to keep generating so many when adjusting the prompt. Thanks for sharing your thoughts!

    • @devnull_
      @devnull_ Місяць тому

      @@CodeCraftersCorner btw thanks for trying these out, I personally simply try to prompt things that didn't or did work with earlier models while swapping samplers and schedulers, so it is nice to see some sort of structured overview too.

  • @krakenunbound
    @krakenunbound Місяць тому

    It would be nice that when you do these videos you would, at the end, flat out say what your favorite settings are for what kind of art. Just tell us what to use.

    • @CodeCraftersCorner
      @CodeCraftersCorner  27 днів тому

      Thanks for the feedback! For me personally, i am using these: sampler: euler, scheduler: sgm_uniform, steps: 20 for (finding the composition), 40 (final image), cfg: 4.5 - 5.5 to make the model follow the prompt.

  • @francaleu7777
    @francaleu7777 Місяць тому

    could you explain what model sampling is?

    • @CodeCraftersCorner
      @CodeCraftersCorner  Місяць тому

      Hello, I'll try. The sampler dictates the noisy image will change to a detail one. The model sampling shift how this happens.

    • @admiralevan
      @admiralevan 19 днів тому

      think of it like a mathematical process that takes noise and interprets it a certain way, you will get different results and completely different final images depending on which one you use with different models. Some are based on similar processes so they have similar outcomes, but you will need to test them all to see what you prefer most

  • @GreenAppelPie
    @GreenAppelPie Місяць тому

    Not impressed. What is 3.5 supposed to offer?

    • @CodeCraftersCorner
      @CodeCraftersCorner  27 днів тому +1

      Hello is a diffusion based model. So we can expect same improvements as with sd1.5, sdxl models.