Join
4
13 Comments

Seeking partner to offer fast StableDiffusion to the impatient

by Jonathan Pelletier

EDIT: Come hang out with us on Discord: https://discord.gg/4s9fH5ft

I love generating images with Stable Diffusion. There is something magical about being able to turn a simple prompt into an image, like the image of this post that I generated for a twitter banner.

I've tried Automatic1111 for image generation at first but found it a bit clunky to be honest. I do not have a GPU so I used to run it on google colab free tier. Since I am confortable with python scripting and the diffusers library, I decided to cut the middle man and use plain google colab notebook + diffusers code.

Even so, I am still annoyed at the slowness of the inference pass for generating a few images on the free tier so I ended up paying for google colab pro.

Now I can access a A100 NVIDIA GPU, but I have to pay a not so cheap monthly fee for it. I also always have to re-install everything and re-download my models everytime I spin up a new runtime and it's midly annoying.

After going through that pain, I discovered what I really want:

  1. Just spin up a terminal and generate images from a model running on a cloud GPU with something like this: selfdiffusion generate --model-id="some_model" --prompt="a person riding a horse" --samples=10

  2. Have those results streamed back to my laptop so I can see them right in my terminal (imgcat sample.png).

  3. Just pay for what I use and not pay for a monthly fee.

I'm looking for someone to help me build this tool I desperately want for myself and I think other people would love it too. These are so many almost hours long youtube video on how to setup this automatic1111 and use it that I feel a simple: pip install selfdiffusion; selfdiffusion generate (...) would be a breath of fresh air for the impatient generative AI hobbyist like me.

There are many more details to the ideas (e.g: offering exclusive models, facilitate fine-tuning, a profit sharing model with model creators etc) but the description so far captures the gist of the idea. Feel free to ask me if you have other questions.

Right now I have an MVP that allows users to sign up and login and I will soon get some inference code running on a serverless inference backend! This is on AWS right now.

Now a bit more about me:

  1. Not employed at the moment so I can partner up full time (40-50 hours a week).
  2. I live in Montreal, Canada.
  3. I have a B.ing in software engineering from Ecole Polytechnique de montreal.
  4. I'm comfortable with my infrastructure skills on AWS and can take care of a scalable backend from code to prod.
  5. I've got no design/UI/Marketing skills but I'm ready to take on any work that needs to be done for success.
  6. I'm mid 30's.

I hope we will connect soon and team up to build something concrete, fast! Also open to partnering up on your idea.

Cheers hackers,

Jonathan Pelletier's avatarJonathan Pelletier
on May 30, 2023
Say something nice to jopela…
Post Comment
  1. 2

    hey, would love to chat. I'm the founder of Catbird

    Jay19917's avatar Jay19917
    ·
    3 years ago
    ·
    Reply
    1. 1

      Come hang out with us on discord: https://discord.gg/4s9fH5ft

      Jonathan Pelletier's avatar Jonathan Pelletier
      ·
      3 years ago
      ·
      Reply
  2. 1

    Hi Jonathan, I'm a "modern fullstack" frontend dev, good complement to your backend skills. I've gotten heavily into AI art, 12K images on MJ so far, explored nearly 20 other AI generators.

    For me, UI/UX is king, and making an image-gen app has rapidly moved up my list of possible projects for this year, despite the competition. In my opinion, excellence will always make a space in the marketplace, even with lots of competition. In other words, I want an app with really excellent UI for making AI images, both as a webapp, and desktop locally. Beyond that, I would certainly want to explore some innovations in models - but there are so many avenues to explore there. My first priority would be a strong UI and operational SaaS. I was already looking at cloud GPU for this purpose, RunPod is currently top of mind for value (I already put in a waitlist request for the new H100 GPUs coming), but that's just a novice perspective at the moment.

    I have been taking a break from development, but I believe in synchronicity and taking opportunities as they arise; your post fits too well with what I'm looking for to ignore. So I hope at least we chat a bit and see if there's a vibe or a vision we can share.

    Send me an email and let me know what platform you want to use to chat, or just reply here and we can connect (Discord, phone, Zoom, Vibehut, Telegram or whatever) hello@mikepeiman.dev

    mountaindad's avatar mountaindad
    ·
    3 years ago
    ·
    Reply
    1. 2

      Founder of Catbird here, we just broke 200K MAUs and 700 paying subscribers, would love to chat

      Jay19917's avatar Jay19917
      ·
      3 years ago
      ·
      Reply
      1. 1

        Yes, me too! Checked out your app - love it.

        We can connect by Zoom, casual and open intro call, schedule a time here if that works for you: https://calendly.com/mikepeiman/collaborative-discussion
        But feel free to email me, hello@mikepeiman.dev and we can get onto a Discord chat, phone call or whatever.

        Thanks for reaching out!

        mountaindad's avatar mountaindad
        ·
        3 years ago
        ·
        Reply
    2. 1

      Just set up a Discord server if you want to reach me that way, https://discord.gg/wk87J2xb (expires in one week)

      mountaindad's avatar mountaindad
      ·
      3 years ago
      ·
      Reply
  3. 1

    Not super sure what the unique part of this service is, wouldn't it be very crowded with services just like it?

    If you just want a couple of images, you could go to: https://huggingface.co/spaces/stabilityai/stable-diffusion

    Not saying you shouldn't try, but I think it might help if you can cater to a specific niche.

    Only slightly related, I probably would be interested to collaborate on AI for interpreting images. I have a mobile wireframing app (TinyUX), and would love to add the feature where you can import an image of a wireframe that is interpreted by the software. So you could immediately update the text, move parts around, etc.

    IdeaGrowr's avatar IdeaGrowr
    ·
    3 years ago
    ·
    Reply
    1. 1

      A use case we want to support is fine-tuning models and then do inference on them. you could use this for example to fine-tune using dreambooth: https://huggingface.co/spaces/multimodalart/dreambooth-training

      However, since this is a web based application, it would not be so straightforward to fine-tune a lot of models. Say you are preparing a birthday party and want to generate a picture of 15 different kids riding a dinosaur given a dataset image of them. With selfdiffusion you could do something that looks like this:

      for data in `ls datasets`
do
   selfdiffusion fine-tune --base "base-model" --dataset $data 
done

      and when training is done, you could generate 100 pictures with each models and see which ones you like better:

      for model in `selfdiffusion list-models`
do
   selfdiffusion generate  \
   --model-id $model \
   --prompt "a picture of a sks person riding a dinausor" 
done

      So the big difference is that: in huggingface, the primitive you play with is access to a GPU, but in selfdiffusion the primitive you play with are a few use cases that you can combine in the way you want. We just so happen to package their execution on a gpu backed runtime for you.

      Second, we will create a private repo of models where creators of model will be able to upload them for other users to do inference. We will share profits with the model creator and thus, we hope that our platform will have models that no other platform has.

      Jonathan Pelletier's avatar Jonathan Pelletier
      ·
      3 years ago
      ·
      Reply
      1. 1

        Thanks for the answer.

        IdeaGrowr's avatar IdeaGrowr
        ·
        3 years ago
        ·
        Reply
  4. 1

    Hey Jonathan, I like your idea!

    I am Igor, atm I am building https://yawplan.com, but kinda need a break. I am mostly about design & frontend, but I have limited backend chops as well (I can spin up a login + payment flow on Supabase+Stripe and iOS\Android purchases in a reasonable time).

    Let's talk on discord (Igor G#8755) or telegram(@gigorr)?

    gigor's avatar gigor
    ·
    3 years ago
    ·
    Reply
    1. 2

      I will definitely reach out to you.

      Jonathan Pelletier's avatar Jonathan Pelletier
      ·
      3 years ago
      ·
      Reply
    2. 1

      Hey Igor, I looked at YawPlan. I like what you're doing and maybe there is potential for more than one collaborator/collaboration. I just replied to Jonathan also as the AI image gen has been a priority for me this season.

      But I like your app, at least at first glance. I used to be in the fitness space, 10-17 years ago, I was a CrossFit instructor, personal trainer and snowboard instructor in Whistler for some years. Health & fitness are/have been a big part of my life. Actually, workout program apps was one of my earliest ideas that drew me into web development, though ironically I've never made anything finished there.

      Anyhow, would be glad to connect with you for possible future synergies and collaboration, given our intersecting interests.

      Just set up a Discord server if you want to reach me that way, https://discord.gg/wk87J2xb (expires in one week)
      or email hello@mikepeiman.dev

      mountaindad's avatar mountaindad
      ·
      3 years ago
      ·
      Reply
  5. 1

    Interesting idea 💡 Kudos and looking forward to your future posts on the product. I always love an idea that provides value, is fast, is easy and is value based pricing, especially usage based pricing and not the usual monthly subscription etc

    karuppiah7890's avatar karuppiah7890
    ·
    3 years ago
    ·
    Reply
    1. 0

      This comment was deleted 3 years ago

      Jonathan Pelletier's avatar Jonathan Pelletier
      ·
      3 years ago