FLOSS AI Tools 🧠

Tired of dystopian, corporate AIs? Check out these Free/Libre Open Source Software AI tools!

Since learning about all these incredible generative AIs and large language models, I've noticed a bit too much mentioning of the names "OpenAI" and "Microsoft" for my liking. Isn't there a less "evil-corporation"-style alternative to them? Less marketing BS? Less branding? For a company called "OpenAI" they sure do share their fair amount of AI with the community, which is basically nothing. Large language models are trained on data from the internet, produced by billions of users. Solely for that reason, AI should belong to the people, after all, they made it possible. In spirit of that, I want to bring some FLOSS tools into the spotlight

Get started

Get a Hugging Face account. Hugging Face is a platform for sharing and inferencing machine learning models. You need this to use some of the tools on the list and to generate an API key in case that a model outside the website needs it. You can obtain such a key by going to your Hugging Face account:

Account ➡ settings ➡ Access Tokens. Keep that tab open! Also get Python and git.

GPT4ALL 🔗

Open source, local free chatbot based on llama by Meta

  • Head to repo, clone it, download model checkpoint (takes a while!)
  • Put the downloaded file in the chat folder and run 
    cd chat;./gpt4all-lora-quantized-OSX-m1 # M1 OSX/Mac
cd chat;./gpt4all-lora-quantized-linux-x86 # Linux
cd chat;./gpt4all-lora-quantized-win64.exe # Windows
cd chat;./gpt4all-lora-quantized-OSX-intel # Intel OSX/Mac

    Chat with your chatbot in the terminal!

Vicuna/FastChat

Also based on llama but fine tuned to achieve a 90% quality rating from GPT4.

  • Visit chat-lmsys for a web application
  • or install locally with pip3 install fschat
  • Launch instructions are in the repo

Stable Diffusion 🔗

Already well known open source text2image tool from Diffusion Labs

  • Run within a Google Colab cell or from a local Jupyter server if you have an NVIDIA GPU!
  • These 11 lines of code are literally all it takes: 
    # Jupyter or Google Colab notebook
!pip install diffusers==0.10.0
!pip install transformers scipy ftfy accelerate
!pip install "ipywidgets>=7,<8"
import torch
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4",
                                           revision="fp16",
                                           torch_dtype=torch.float16)
pipe = pipe.to("cuda")
org_prompt "the best website known to humanity"
image = pipe(org_prompt).images[0]
image.save("output.png")
image

I will prompt it with the best website known to humanity and oh boy did it deliver!

Modelscope Text2Video 🔗

Run this text2video generator directly in HuggingFace or in a Google Colab notebook!

Prompt: A can of soup flying through space Andy Warhol style. As you can see, it kind of skipped the space part but the Andy Warhol is definitely there!

Uberduck 🔗

Text2Voice for your favorite voices!

  • Create an account, select a universe, select a voice and type your text
  • You will now hear a message from family man and father of the year award winner Walter White

Riffusion 🔗

Browser based Text2Music tool.
Since I want to stay hip and young I will ask this tool for some good tunes that the kids like these days. Prompt it with anything and it will try to transform it into a beat!

while you're listening to some quality music you can also try:

Segment anything (Meta) 🔗

Context understanding rotoscoping tool:

  • gone are the days of manual or color based rotoscoping
  • this tool understands context and layers
  • you can select single "components" in your image and open them as different images on external tabs
  • This makes image manipulation so much easier

PUBLIC SERVICE ANNOUNCEMENT

If you're just interested in AI tools, skip this part.
You might ask yourself: Why should I use these lesser, more complicated tools with weird design when I have ChatGPT and Microsoft Copilot at my side? Let's gain some distance and explore that question:

  • you and me are part of humanity and probably not all too mighty or influential, otherwise you wouldn't be here
  • You and others fill the internet with content: your comments, search histories, forum posts, images and on and on
  • This data, without your consent, was used to train these large language models. Without millions of internet users, you included, they wouldn't exist
  • By pouring more and more money into one company, they will be able to monopolize on that technology, moderating its abilities to their wish, creating their version of reality
  • They will most likely obey the highest bidder, amplifying their incentives and muting others
  • Essentially, this will create a pay to win system for any form of AI applications, including politics, market trends, advertising and journalism
  • You might realize that the highest bidder does not essentially wish for the highest degree of truth
  • Meaning that he who is on top of the AI industry, controls information flow in an age where human-created works and AI content become more and more indistinguishable from another
  • The only way this can be avoided is by distributing these systems, creating alternatives and options for a healthy AI ecosystem
  • These models can have their paid services for companies and ease of use, but the core, the architecture and weights, need to be opened up to the public so that we can understand them and choose with what kind of AI we want to interact with
  • Each person partaking in the usage or development of FLOSS models hinders companies from achieving a monopoly a little bit and helps open models to grow
  • And if that means that I get the top notch AI tools a week or two later than the people who pay for it, while helping an open AI platform to bring out the best in everyone for everyone then i'll gladly wait for them. Choose your side.

BONUS: tools that aren't open source but FREE:

Blockade Labs 🔗

Text to 3D landscape texture tool.

  • Type in your prompt and find a world even better than you could think of
  • select a style
  • generate and be prepared for your RAM being eaten by your browser
  • As you can see, the soup got a slight refurbishment along the way
  • You can also download this 360 degrees image and use it for concept backgrounds in your games and animations

if you're not happy with the images, try

Latent Labs 🔗

Image2Image Reimagination:

  • Describe a style
  • Select an image
  • I will select my YouTube avatar, also generated by stable diffusion, to get a more cinematic look

Not quite what I asked for but still really cool! Since I mentioned stable diffusion again, try out

Diffusion Art 🔗

Web based conglomerate of different text2image AIs

  • Choose your AI
  • Write a prompt

It seems that next time, I will consult an actual webdesign AI instead

By the way, the most helpful website for my research was futuretools.io, a search engine for new and future AI developments. Check that out for yourself!