• davely 8 days ago
    I've been working on a web client[1] that interacts with a neat project called Stable Horde[2] to create a distributed cluster of GPUs that run Stable Diffusion. Just added support for SD 2.0:

    [1] https://tinybots.net/artbot?model=stable_diffusion_2.0

    [2] https://stablehorde.net/

    • davidkunz 8 days ago
      Wow, this a great site, thanks for the links!
  • bryced 8 days ago
    Try out the pre-release like this:

    `pip install imaginairy==6.0.0a0 --upgrade`

    New 512x512 model supported with all samplers and inpainting

    New 768x768 model supported with the DDIM sampler only

    Not yet supported is the upscaling and depth maps.

    To be honest I'm not sure the new model produces better images but maybe they will release some improved models in the future now that they have the pipeline open.

    • swyx 8 days ago
      congrats! how did you upgrade it so fast? and what would you call out as the main technical pointers to adapting the base release for M1's?
      • bryced 8 days ago
        All the same issues as migrating 1.5 to M1s. It went fast because I upgraded my existing codebase that had those fixes already instead of building of the new compvis one.
  • greggh 8 days ago
    This is awesome, but I still like using the GUI for m1/m2 Macs, DiffusionBee.


    • malshe 8 days ago
      Thanks for sharing this. I was looking for something simple like this
    • diebeforei485 8 days ago
      Does this use Stable Diffusion 2.0?
    • jibbers 8 days ago
      And apparently Intel Macs also! I had no idea!
  • Smaug123 8 days ago
    Nicely done; this seems to work for me. In my own attempt, I got stock Stable Diffusion 2.0 "working" on M1 using the GPU but it's producing some of the most cursed (and low-res) images I've ever seen, so I've definitely got it wrong somewhere. The reader can infer the usual rant about dynamic typing causing runtime misconfiguration in Python.
  • typest 8 days ago
    How much of this is stable diffusion 2, and how much is something else? For instance, the text based masks, the syntax like AND and OR, the face up scaling — are these all part of stable diffusion 2 (and can be used via other stable diffusion apis)?
    • bryced 8 days ago
      - text-based masks use a clipseg model. - the boolean mask logic is unique to this library - the face fixing is done by CodeFormer
  • fareesh 8 days ago
    What's the minimum VRAM requirement?
  • yreg 8 days ago
    As with previous macOS Stable Diffusion tools, this is Apple Silicon only.
    • smoldesu 8 days ago
      If you have an Intel Mac with sufficient memory, it's totally possible to run it on-CPU as well.
      • dylan604 8 days ago
        >If you have an Intel Mac with sufficient memory,

        which means what? why be so ambiguous. If if needs 16GB, say so. If it needs 32, say so. your sufficient memory comment is insufficient

        • smoldesu 8 days ago
          The figure isn't static. Some models require as little as 3.5gb of free memory, others demand 8-16 gigs. MacOS is weird with memory management and everyone's Mac is different; I'd really only recommend running the model on 32-gig machines to avoid writing into swap, but technically it's possible with 8 and 16 gig machines.
  • gbighin 8 days ago

    > A decent computer with either a CUDA supported graphics card or M1 processor.

    Why so? How does an M1 processor replace CUDA in a way a x86_64 processor can't? Do they use ARM assembly?

    • pavlov 8 days ago
      It’s not the ARM core but the integrated GPU in the M1. It has access to the entire main memory unlike a traditional GPU with its own local VRAM.
      • gbighin 8 days ago
        Oh, interesting! But does it support CUDA? How is the integrated GPU used for ML tasks?
        • hnarayanan 8 days ago
          Both PyTorch and TensorFlow offer backends for Metal that works pretty well on Apple Silicon.
        • dagmx 8 days ago
          To add to what people said, most of these ML models target an ML library like TensorFlow or PyTorch.

          Those in turn have hardware accelerated backends. Traditionally they’ve only had CUDA backends but Apple ported large chunks of both to Metal as well.

          So none of these libraries really target CUDA. In fact they’d run fine without a supported GPU but much slower.

        • Filligree 8 days ago
          It does not support CUDA; SD does not require CUDA.
        • pavlov 8 days ago
          I believe there’s a Tensorflow acceleration adapter for Apple’s ML API which uses Metal behind the scenes.
        • malshe 8 days ago
          pytorch can use the GPUs on M1 macs. Sebastian Raschka's post explains it nicely and shows some benchmarks too. https://sebastianraschka.com/blog/2022/pytorch-m1-gpu.html

          From his post:

            if you want to run PyTorch code on the GPU, use torch.device("mps") analogous to torch.device("cuda") on an Nvidia GPU.
          • crucialfelix 8 days ago
            In some cases there are operations not supported on mps. For those set:

            os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"

            and it will run on cpu if some operation isn't supported

            • malshe 8 days ago
              Excellent! Thanks
  • anothernewdude 7 days ago
    2.0 is a mixed bag. It's set making pixel art back entirely. I'm pretty sure this is down to the aesthetic filter - it has a very biased idea of what good images are. It's silly to do that at the training stage, that should be something you do in the prompt.

    Fine tuning is out of reach for me, so I'm sticking to 1.5.

  • lostintangent 8 days ago
    Wow, this looks awesome! I noticed that the sample notebook doesn’t include SD 2.0 by default, and says that it’s too big for Colab. Is that a disk size/RAM limitation?

    As an aside, it would be cool if you versioned that notebook in the repo, so that it could be easily opened with Codespaces.

    • bryced 8 days ago
      Yeah I tried to get it running but it kept crashing with "out-of-ram" errors.

      Good idea to version the notebook.

  • egeozcan 8 days ago
    This would have been perfect if it worked on Windows too. I need to look into dual booting Linux (opening a can of worms) just to give it a try, as WSL doesn't seem to cut it.
    • bryced 8 days ago
      It might work on windows but I haven't tested it there.
      • patates 8 days ago
        It only uses the CPU. Somehow the GPU detection fails.
      • dekhn 8 days ago
        for me the pip install on windows (anaconda) failed installing basicsr: error: metadata-generation-failed
        • bryced 8 days ago
          I don't think it works with anaconda on any OS.
    • satvikpendem 8 days ago
      Why not use Automatic1111's? I think he already added SD 2.0.
    • boycott-israel 7 days ago
      fwiw dual booting is ultimately simpler than WSL and it's quirks
  • underlines 8 days ago
    is it possible to add volta or xformers for a massive speed increase?


    • bryced 8 days ago
      Possibly. Haven't tried. In principle should be possible.
  • superpope99 8 days ago
    This seems to work for me. Incredible work turning this around so quickly!
    • habibur 8 days ago
      If you are running it natively [ not on a cloud ] what's the ram size of your graphics card?
  • semicolon_storm 8 days ago
    Pretty slick, SD 2.0 performance actually seems to be better than 1.5?
    • bryced 8 days ago
      You're probably noticing the newest sampler, which also works with 1.5.
  • algon33 8 days ago
    Nice, a friend was looking for something like this.
  • TekMol 8 days ago
    What is a good VM to try this out?

    Something on AWS, Hetzner etc?

    • petercooper 8 days ago
      AWS g5.xlarge instances. Very fast (roughly RTX 3080 speeds) and about $1 an hour. However, you can just turn the instance on and off and not pay anything except the latent EBS cost.
  • 88stacks 8 days ago
    awesome library, I haven't seen this before. I just added it to my stable diffusion api service so you can query stable diffusion 2.0 if you don't GPUs setup currently: https://88stacks.com
    • ttpphd 8 days ago
      Why is it called 88 stacks?
      • turnsout 8 days ago
        Also wondering about the 88—only because of its Neo-Nazi/hate-speech connotations