Stable Cascade Comfy Checkpoint Update

แชร์
ฝัง
  • เผยแพร่เมื่อ 5 พ.ย. 2024

ความคิดเห็น • 30

  • @garrettdonahue2394
    @garrettdonahue2394 8 หลายเดือนก่อน +2

    I would love to look at that multi-workflow. Great stuff!

    • @FiveBelowFiveUK
      @FiveBelowFiveUK  8 หลายเดือนก่อน +2

      I released the Faceswap version today that used Reactor to improve faces to a target.
      I'll get the big workflow out in a few days, gotta add more changes for a video :)

  • @WanerRodrigues
    @WanerRodrigues 8 หลายเดือนก่อน +2

    Really cool! Is nice you being chasing the newer versions, I appreciated that! It seams the compression value has a relation to the quality and resolution, seams 64 are good for really large imagens, otherwise the 42 seams better to 1024 images

    • @FiveBelowFiveUK
      @FiveBelowFiveUK  8 หลายเดือนก่อน +2

      yes one thing i found today that i had not mentioned was at 3072x3072 you can get good results with 72 compression. That makes me think that "divisible by 8" rules applies and the value will scale to resolution. but it's hard to make concrete rules as always.

    • @WanerRodrigues
      @WanerRodrigues 8 หลายเดือนก่อน

      @@FiveBelowFiveUK That's make a lot of sense! I will try it, thank you

  • @sprinteroptions9490
    @sprinteroptions9490 8 หลายเดือนก่อน +1

    very much appreciated. Works nicely w/ new comfy models.. very impressive actuallty on 4gb vram and 20? steps total? wow. game changer i hope

    • @FiveBelowFiveUK
      @FiveBelowFiveUK  8 หลายเดือนก่อน +1

      thanks for that feedback, running on 4GB is crazy!

    • @springheeledjackofthegurdi2117
      @springheeledjackofthegurdi2117 8 หลายเดือนก่อน

      what was your generation time/performance like? I ask due to having a 6gb here

    • @sprinteroptions9490
      @sprinteroptions9490 8 หลายเดือนก่อน +1

      @@springheeledjackofthegurdi2117 Hi! After the comfyui models came out i can do a 1024 size image w 20 steps in 60 seconds w/ his v30 argus workflow. I can even do his latest v82-Cascade-Anyone workflow in about 200 seconds!

  • @glenyoung1809
    @glenyoung1809 8 หลายเดือนก่อน

    If there is anything I've learned, it's that patience pays off and it's not always an advantage to be an early adopter!
    The new comfyui checkpoints are very efficient memory-wise.
    The largest amount of VRAM used when loading and running the workflow, was 8.4GB for the models only not counting any other system overhead in VRAM(other apps such as YT etc).
    Comfy used 16GB of system RAM to cache the models and other data.
    So overall these changes are much more efficient.
    Those with 8GB cards should be able to run these workflows but they'll get a bit of "swapping" where Comfy will need to access system RAM for some of the model data but it shouldn't crash.
    Not sure what will happen with much smaller VRAM cards like 6GB which might bog down with memory transfers between the GPU and system RAM.

    • @FiveBelowFiveUK
      @FiveBelowFiveUK  8 หลายเดือนก่อน +1

      overall this is the most promising model i've seen since 1.5, it exceeds expectations on al fronts. If we can get training and controlnets up and running, it's a done deal now.

  • @guns1inger
    @guns1inger 8 หลายเดือนก่อน +1

    Great work. Been looking at an extension for Forge, but it takes 5 minutes an image, versus 20 seconds in comfyUI, and the comfy images are way bigger.

    • @FiveBelowFiveUK
      @FiveBelowFiveUK  8 หลายเดือนก่อน

      wow that's a big stat. not a small difference there!

  • @stevietee3878
    @stevietee3878 8 หลายเดือนก่อน

    Fantastic stuff mate !

  • @djivanoff13
    @djivanoff13 8 หลายเดือนก่อน

    you're the best!please add a Controlnet node

    • @FiveBelowFiveUK
      @FiveBelowFiveUK  8 หลายเดือนก่อน +1

      afaik the canny model support is not working yet, but it will be added to a workflow and then a video will come as soon as it lands!

  • @JalenJohnson1234
    @JalenJohnson1234 6 หลายเดือนก่อน

    I noticed the file sizes of the checkpoints are lower than the Unet models. Are the checkpoints less quality than the previous unet models?

    • @FiveBelowFiveUK
      @FiveBelowFiveUK  6 หลายเดือนก่อน

      strangely the checkpoints have included several other files, so it's likely a combination of the smaller versions repackaged. Stage B has the VAE encoder and clip, while stage C included the VAE decoder. It's a new architecture with some interesting strengths and unfortunate weaknesses, however someone told me to try training Cascade Lora on 1536x1536 dataset, to resolve the output texture issues. I have not had time to investigate this due to new released models that we had to look at first.

  • @MarkNelson-zo1lr
    @MarkNelson-zo1lr 8 หลายเดือนก่อน

    Great stuff, but I'm having issues getting Reactor node installed, Comfi just doesn't seem to like it...

    • @FiveBelowFiveUK
      @FiveBelowFiveUK  8 หลายเดือนก่อน

      you might be able to run "pip install -r requirements.txt" inside the /custom_node/folder for Reactor. Also there is an Inswapper8.onnx model that is in the comfy manager models list. I'll do a video on the new stuff (since this video) next

  • @HasanAslan
    @HasanAslan 8 หลายเดือนก่อน

    I am on amd and always have memory problems with models higher then sd1.5 (recently solved it by using lowvram always with sdxl) and cascade seemed like it worked without any problems at all with the previous 3 file method. Now that I see the two models are big -one of them 9 gb- no way I could run them successfully without massive slowdowns etc. Do we have to switch to these ? Are there any advantages of using these over 3 seperate files ?

    • @FiveBelowFiveUK
      @FiveBelowFiveUK  8 หลายเดือนก่อน

      the FP16 lite version are for Ampere cards (nvidia 40xx/30xx) so you might be able to use the older workflows with stage B&C lite. The big advantage of the newer checkpoint method is portability, each checkpoint also carries the text encoder, CLIP Vision and a VAE.
      What i'm finding is that unless you have nvidia 30xx/40xx and or under under 12GB definitely under 6GB is better off using my Cascade workflows from Argus-v12 to v18.
      When i shared them someone from stability told me they were outdated (same day it all came out haha)
      TLDR is, if it worked for you before and not with the checkpoint method, stick the the original. I'll release some updated versions of those too because so many people are still using them.

  • @yklandares
    @yklandares 8 หลายเดือนก่อน +1

    эта ошибка на всех 4 рабочих поцессах выдает

    • @97BuckeyeGuy
      @97BuckeyeGuy 8 หลายเดือนก่อน

      Slava Ukraini!

  • @DonNardi
    @DonNardi 8 หลายเดือนก่อน

    I loaded the checkpoints correctly but I get this following error "Error occurred when executing CLIPTextEncode: 'NoneType' object has no attribute 'tokenize'"? My gpu is a 2070 Super. Not sure whats causing this error?

    • @FiveBelowFiveUK
      @FiveBelowFiveUK  8 หลายเดือนก่อน

      ok this is a strange one, The Clip encoder should be built into the Stage C Checkpoint. check that the positive and negative prompt text boxes are being connected to the chackpoint loader CLIP for Stage C. let me know if that helped.
      I had to reupload one of the workflows which was missing a node connection, it's possible you took that copy before the fix?

    • @DonNardi
      @DonNardi 8 หลายเดือนก่อน

      I got it to work by updating Comfyui!
      Thanks for your reply.

  • @ДиДи-м3ю
    @ДиДи-м3ю 8 หลายเดือนก่อน

    does not work

    • @FiveBelowFiveUK
      @FiveBelowFiveUK  8 หลายเดือนก่อน

      well there are a lot of things that could cause that. We assumed you installed python 3.10.9 and Torch 2 with Cuda 11.8. This can be quite a challenge for people. In addition you have two methods to use, this one that used two checkpoints and for those with less VRAM on GPU, you have a manual method with loads everything manually. With thousands of people reporting success across many specs, i cannot guarantee much but this definitely works. Good luck ;) because if it's not working, you likely just need to try and follow the instructions again.