ComfyUI - Vast.ai: tutorial - how to rent cheap high VRAM GPUs for your AI art

Поділитися
Вставка

КОМЕНТАРІ • 47

  • @brgtubedev001
    @brgtubedev001 9 місяців тому +2

    Thank you for showing us how to set up this method of cloud computing. You have an excellent way of explaining the method. Good luck to you and your channel!

    • @koalanation
      @koalanation  9 місяців тому

      Thanks to you for watching!

  • @victorrochin2616
    @victorrochin2616 10 місяців тому +2

    Just what I needed! Once that I need to delete my instance and pick up my work the next day, is there a way of saving my data and pick up where I left later in another instance?

    • @koalanation
      @koalanation  10 місяців тому +1

      You need to synchronise with Dropbox or Blackblaze. Last time I checked with Google Drive did not work. You can always stop the instance, but you have to pay full price, unfortunately.
      Alternatively, you can transfer files from a local computer via SSH or use another cheaper instance for storage, but then you depend on the owner of the instance.

  • @TheFusssi
    @TheFusssi 5 місяців тому

    Thank you :) I got it working with the pytorch version!

    • @koalanation
      @koalanation  5 місяців тому

      Excellent! Have fun

    • @TheFusssi
      @TheFusssi 5 місяців тому

      @@koalanation short catch-up -> used it multiple times now + adapted your script and it works like a charm. Thank you very much for the video, as without it I would either have spended way more time setting it up or even have given it up before getting it to work.

    • @koalanation
      @koalanation  5 місяців тому

      @@TheFusssi good to hear!

  • @marcosansalone
    @marcosansalone 13 днів тому

    Great tutorial. I see Docker so I assume is running on ubuntu, correct?

  • @alexlindgren1
    @alexlindgren1 Рік тому +1

    Thanks for this! I want to set up a ComfyUI API instead, so I can send a request from another server and retrieve the generated images, how would you recommend me doing this?

    • @koalanation
      @koalanation  Рік тому

      This is a little beyond my expertise. I have seen a tutorial for stable diffusion A1111 with endpoints for runpod. Maybe you want to look at it to get some ideas how to do it for comfyui.

  • @ceryx6849
    @ceryx6849 2 місяці тому

    Alright i wanted to know if non knowledgable everyday user could use this and it is not the case. I thought it was just about clicking, paying and then you get an additional virtual GPU on your computer. Thanks.

    • @koalanation
      @koalanation  2 місяці тому

      Hi! You can maybe you can check out the SwarmUI templates. I am preparing a runpod tutorial, in where the basically the script is in the template. So not 'typing' needed for the installation. However, starting comfyui is still from a terminal.
      Alternatively, services like RunDiffusion or ThinkDiffusion offer a very easy to use interface for ComfyUI or other SD webui's

  • @Vinny_views
    @Vinny_views Рік тому +1

    Stable diffusion fast is missing from the template

  • @prasanthrajappan4401
    @prasanthrajappan4401 Рік тому +1

    how can i have an api to this set up to access it externally

    • @koalanation
      @koalanation  Рік тому

      Hi! I am not a technical person, not sure of what your question is. In vast.ai there is some documentation about API. I understand access to the VM can be done via SSH/SCP but I just simply download the requirements via git clone and wget, and use jupyterlab to upload/download from my local machine . Access to Google/drive and Dropbox is possible, but I have not used it. Sorry I cannot help more.

  • @EchoesofHeartbreak
    @EchoesofHeartbreak Рік тому +1

    owesome :)

  • @yasinaydn763
    @yasinaydn763 8 місяців тому +1

    can you share scripts again? pastebin is not working.

    • @koalanation
      @koalanation  8 місяців тому

      Hi! It works for me...🤔 Anyways, I think I may need to update it. To begin, you can use more general templates, new versions of pytorch, etc...I think it is also valuable to connect to have the models/files in backblaze and sync it, instead of downloading everything every time. Takes the same time, but at least you keep the work....

  • @Scramit
    @Scramit Рік тому

    why i don't have manager in your interface?

    • @koalanation
      @koalanation  Рік тому

      In the script the manager should be installed. Anyway, follow the instructions in the GitHub:
      From the ComfyUI folder
      1- cd custom_nodes
      2- git clone github.com/ltdrdata/ComfyUI-Manager.git
      3- Restart ComfyUI

  • @congtaihu1287
    @congtaihu1287 9 місяців тому

    i could not find the template you are referring to

    • @koalanation
      @koalanation  9 місяців тому

      Use the pytorch template

  • @momoe7563
    @momoe7563 Рік тому

    Great tutorial, do you think it's possible to edit the template for it to automatically install comfyui with the user going into Jupiter?

    • @koalanation
      @koalanation  Рік тому

      Thanks!
      I think it should be possible to make a jupyter notebook from the script. I tried at some moment but I had some issues and decided it was faster to directly just copy and paste the script in the terminal. I have the code in a Google Docs document, which I use every time I want to use vast.ai

  • @Mnbir
    @Mnbir 4 місяці тому

    Greate Video, Thanks

  • @AIWarper
    @AIWarper Рік тому

    Do you have to do this literally every time you launch? Also, there is no "fast" template anymore

    • @koalanation
      @koalanation  Рік тому +1

      If you destroy your instance at the end of the work to keep the costs low, yes. Alternatives are to using your cloud storage or to create your own docker container with the basic models and nodes you normally use.
      Instead of the fast template, you can use the pytorch template.

    • @AIWarper
      @AIWarper Рік тому

      @@koalanation how would I go about the latter (creating the docker)?

    • @koalanation
      @koalanation  Рік тому

      @@AIWarper if you are using a 'base' of models and nodes and not changing many things, I think it should not be very difficult to create an image, if you know how to use docker. I have not done it myself, though.
      If you like to try things and you have good reference workflows, I think the fastest is to only install ComfyUI and ComfyUI manager, then use the install missing nodes feature from the manager, and install the models with the manager too. But yes, the cheapest solution may not be optimal if you want not to run these scripts anytime.

  • @Mindset2Work
    @Mindset2Work Рік тому

    Is the choise of the preprocessor / cpu important?

    • @koalanation
      @koalanation  Рік тому

      To be frank, I am not sure. I try to use what is available and so far my needs are covered

  • @saberkz
    @saberkz 3 місяці тому

    Do i need to install the plugin and all each time i run it

    • @koalanation
      @koalanation  3 місяці тому

      Unfortunately, using vast.ai, you need to run the script everytime. With other services, like runpod, you can have permanent storage and then you do not need to download all nodes and models everytime.

    • @saberkz
      @saberkz 3 місяці тому

      @koalanation what about google colab ? Whos better

    • @koalanation
      @koalanation  3 місяці тому

      @@saberkz with Collab I think the best thing is you can easily connect and mount Google drive. For the price others are typically cheaper.

  • @2DPTuts
    @2DPTuts 7 місяців тому

    i wonder if its possible to get this to run on krita

    • @koalanation
      @koalanation  7 місяців тому +1

      There is a guide with a Krita template for runpod or vast.ai: github.com/Acly/krita-ai-diffusion/wiki/Cloud-GPU
      Never tried, though