ComfyUI - Prompt Engineering with CFG, Sampler Steps and Clip Skipping - for Stable Diffusion Users

Поділитися
Вставка
  • Опубліковано 6 сер 2024
  • This video explores some little explored but extremely important ideas in working with Stable Diffusion - at the end of the lecture you will understand the relationship between CFG, Sampler steps and Clip skipping and the negative and positive implications. But take your time with this one, it goes into professional level detail on the subject.
    These are core ideas in helping you to refine your prompts and understanding the methods currently used in Stable Diffusion.
    The previous lectures mentioned in the video can be found in the beginner's guide to ComfyUI
    Course Discounts
    BEGINNER'S Stable Diffusion COMFYUI and SDXL Guide
    bit.ly/GENSTART - USE CODE GENSTART
    ADVANCED Stable Diffusion COMFYUI and SDXL
    bit.ly/RESTAD - USE CODE RESTAD
    Mastery Course for Stable Diffusion (commences with this tutorial)
    cutt.ly/NwcMf3w3
    This lecture also looks at how core techniques previously examined on the course can be applied holistically to improve overall workflow.
    Join 🏆 this channel to get access to exclusive content and perks:
    bit.ly/XOVjoin
    00:00 Printing as a Metaphor for Stable Diffusion
    02:50 Understanding Latent Space, Clip Networks, Diffusers and VAEs
    04:55 Sampler Steps, CFG and Clip Skips
    05:33 The relationship between sampler steps and the checkpoint
    12:40 The relationship between CFG and the Text Prompt
    18:35 The relationship between color, contrast and CFG
    29:00 The CFG is broken
    30:09 High steps number and the relationship with the prompt
    32:56 Clip Skip and the Clip Neural Network
    35:20 The CFG is completely broken
    36:00 Putting this analysis of the neural networks together
    39:40 Possible solutions to fix the broken CFG
  • Наука та технологія

КОМЕНТАРІ • 25

  • @Pixovert
    @Pixovert  11 місяців тому +4

    The previous lectures mentioned in the video can be found in the beginner's guide to ComfyUI
    Take. Your. Time. With This One, it goes into *_professional level_* detail on the subject.
    Course Discounts
    BEGINNER'S Stable Diffusion COMFYUI and SDXL Guide
    bit.ly/GENSTART - USE CODE GENSTART
    ADVANCED Stable Diffusion COMFYUI and SDXL
    bit.ly/RESTAD - USE CODE RESTAD

  • @jaredbeiswenger3766
    @jaredbeiswenger3766 11 місяців тому +4

    Wonderful to listen to. Would love more content like this with other values that we tweak without really knowing what they do. I do a lot of this experimentation on my own, but to be able to watch a summarized version of someone else doing it while I'm working.... that let's me focus the little time I have for my hobbies.

  • @OutOfWards
    @OutOfWards 5 місяців тому

    Thank you for taking the time to make these super helpful videos! You are a God send!

  • @edenorion
    @edenorion 10 місяців тому

    One of the most important explanation that I've heard. Thank you very much

  • @Mehdi0montahw
    @Mehdi0montahw 9 місяців тому +1

    Your way of explaining is wonderful, with details that I have not seen in any video, and the elegant sequence is beautiful, and there is no doubt that you deserve to press subscribe and like.

  • @zraieee
    @zraieee 10 місяців тому

    Well Done, Thanks

  • @SebAnt
    @SebAnt 8 місяців тому +1

    Brilliant!!!!! ❤

  • @0A01amir
    @0A01amir 11 місяців тому +1

    Thank you.
    Really liked efficiency nodes but it's loader (best part of it) is not using lowvram and xformers and resulting in slower generations in low end machines.

    • @OutOfWards
      @OutOfWards 5 місяців тому +1

      open up the bat. Mine looks like this...
      .\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --lowvram
      pause

    • @0A01amir
      @0A01amir 5 місяців тому

      @@OutOfWards Oh comfyui is ok but when i load efficiency node that happens but it won't anymore since i upgraded my GPU two months ago.

  • @artsbybrian_
    @artsbybrian_ 10 місяців тому

    Great video! Mind if I ask for the workflow? Wanna test it.

  • @lioncrud9096
    @lioncrud9096 11 місяців тому +2

    thank you very much, do you have any resources on advance prompting, like wildcard prompts {|} etc, i'm trying to figure out if ComfyUI has a version of alternating "to" and "from" prompts like Auto1111's [man:dog:0.5].

    • @Pixovert
      @Pixovert  11 місяців тому +1

      This is included on the course described above - the beginner's course. ComfyUI is different in its application of some of these methods and techniques

    • @97BuckeyeGuy
      @97BuckeyeGuy 6 місяців тому

      @@Pixovert Great non-answer.

    • @97BuckeyeGuy
      @97BuckeyeGuy 6 місяців тому

      @lioncrud9096 Did you ever, by chance, find some answers on this? I've been looking for a way to do this in Comfy, as well. I can't believe the developer just chooses to not support it.

    • @lioncrud9096
      @lioncrud9096 6 місяців тому

      I did not sadly, basically my understanding is that the alternating prompt format does not work in ComfyUI. The way the prompt is processed in Comfy is different than A111 so some of the syntax don't work in both -_-@@97BuckeyeGuy

  • @spiralofhope
    @spiralofhope 8 місяців тому

    how do I clip skip without the efficiency loader? I like to work without custom nodes.

  • @formixcode
    @formixcode 10 місяців тому

    how to use strength line wire? thx

  • @lioncrud9096
    @lioncrud9096 11 місяців тому

    is there a reason you didn't use the Clip output?

  • @mithrandir404
    @mithrandir404 6 місяців тому

    From a camera perspective. Cfg is the ISO. Steps is shutter speed, and Clip skip is the Aperture

  • @Vestu
    @Vestu 10 місяців тому

    99% of the time CFG should be 5-8

  • @maxpaynestory
    @maxpaynestory 8 місяців тому +1

    "beautfil scenery nature glass bottle landscape, ROUND." Not an actual requirements of any scene. We don't need examples of random prompts generating random images. All we could find on youtube is Comfy UI generating random pictures and videos using random prompts. But infact this tech is not mature enough to be used for creating any scene.
    Try create a scene containing 5 teenage kids playing Call of Duty 2 on their computers, they are happy and cheering up.