AI as a toy

I began derailing another thread conversing with Aravisian about AI, and given that I have a question about a problem I've run into several times anyway, why not start a thread for more general conversation?

There are numerous ethical and societal questions, surrounding AI's use; they're hard to escape online right now. This thread is not for that. All else aside, if you use AI for anything fun, what's your game?

I use AI strictly locally these days, stuff I can run on Zorin. Mainly I poke at LLMs seeing just how well they handle weird prompts (usually terribly, but sometimes amusingly), and generate images with Automatic1111 or Forge and a wide variety of StableDiffusion/XL or Flux.1 models, mostly just for images to amuse myself, rarely for wallpapers, sometimes for art I'll only use temporarily (a character in a tabletop RPG for example), or less frequently, to convey an image in my head with some precision—such as if I mean to actually PAY an artist, but don't have reference images.

(Stuff below is all clean, but hidden to reduce vertical scroll.)

Pathfinder character

Character from a Pathfinder tabletop game. Strix ancestry, her feathered cloak is her actual wings—disguising that way requires a feat, but she was all about lies. Made with NovelAI before I switched to AI tools I can run locally on Zorin.

Bagpipe

Bagpipe from a mobile game called Arknights, which I also play on Zorin, via Genymotion. A farmer/soldier (...the forum dislikes a certain Latin loanword), she mentions potatoes in a dialogue line and the player base kind of attached that to her. The model is specifically designed to make something look like a PVC figure.

Tools I use:
Forge, a fork of Automatic1111 with more features and model support. Setup is as simplest from the terminal, but needs Python 3.10 and git, both easily installed with APT.

Setup
  1. Navigate to where you want to install.
  2. git clone https://github.com/lllyasviel/stable-diffusion-webui-forge.git
  3. Navigate to the cloned directory, should be stable-diffusion-webui-forge
  4. Python3.10 -m venv venv
  5. ./webui.sh

That'll install pre-requisites in the venv (virtual environment, to keep from polluting your actual Python installation with pip based installs) and start a browser window. Setting up models and using it is a bit involved; this is how I started—just skip the Windows install section. Officially, Forge is experimental and the author recommends using Automatic1111, but Automatic1111 doesn't support as many model types, tends to be slower, etc.

Automatic1111, the simplest "at home" tool for running Stable Diffusion based models. Supports Stable Diffusion 1, 1.5, SDXL, and I think Stable Diffusion 2, but hardly anyone adopted 2. Like Forge, it needs Python and is pulled with git. Installation will pull Stable Diffusion 1.5 automatically so you have something to work with, but at it's base, that's not great.

Setup

Exactly the same as Forge above, but clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git instead. As noted, Automatic1111 supports fewer model types and is generally pretty slow to adopt new features these days, but the Forge developer recommends it for stability.

I've had Comfy UI working on Zorin too, but I never really figured it out well and can't find good Linux setup instructions. It's the most powerful and not nearly as comfy as the name suggests, in my limited experience.

Finding more models isn't hard. Civitai.com has tons, mature hidden by default, with options in user profiles to hide anime, furry, gore, and political content, whether it's mature or not.

LMStudio, an all in one for both playing with and developing for LLMs. Its Linux version is an AppImage, just download and run. It has a built in model search that'll let you hunt for and download models from Huggingface without visiting the site, and will let you know which models you're likely able to offload to a GPU, partially or fully.

Finally, my problem: If you've ever used Automatic1111 or Forge, have you been able to install Nvidia's TensorRT extension? It always goes into an infinite "working" animation, but I'm not aware of any reason it shouldn't work on Linux. It also breaks the venv and I have to repair the installation or reinstall.

I'll be honest, I've dabbled with a few AI things a couple times here and there, but for the most part, other than maybe some images and whatnot, I don't use them at all. And even then, I prefer trying to find someon who actually makes a genuine piece of art (or a nice photograph) over and AI generated one. Not that I hate AI photos or anything, just that usually an AI photo has just a bit of .... ethereal-ness to it that I can't get over.

I do kind of like using an AI to try to creat a text based adventure game. Not many people, if any, make any new ones nowadays, and doing that can be fun, until it inevitably has no idea what's going on and the world goes down the drain because it can't keep track of anything.

That being said, I do actually find myself using Co-pilot a lot at work. Not for anything actually work related, but because basic search engines are so useless, that sometimes that AI can just sift throught the garbage and get me the links that I'm actually looking for. SOMETIMES, it is still very prone to being extremely wrong or just not understanding what I''m asking it at the best of times.

The "etherealness" is getting better, though my examples above aren't the best for that. That said, I agree, which is why I mentioned using AI to springboard a conversation with an artist, and using AI images for amusement or temporary purposes.

Definitely fun, but yeah, my experience is the same. I'll amuse myself for a bit, then it'll go to hell in a handbasket. One trick for staving that off, at least with LMStudio (I added a link to that after you replied) is to add details that need keeping to the system prompt, since they don't roll out of the context. It's still far from perfect, but it helps. I've also seeded a second conversation with an LLM by putting notes from the first in the system prompt and making the first prompt a recap. If you're set to keep the first, recent, and truncate the middle, that can work decently.

...we recently had someone ask Copilot to generate a banner image for their team at work, and got someone's social studies homework instead, somehow. I haven't touched it at all.

...we recently had someone ask Copilot to generate a banner image for their team at work, and got someone's social studies homework instead, somehow. I haven't touched it at all.

That's why I don't use it for anything actually important. Just try to use it as a search engine that maybe kind of works in getting me a link or an article that is actually pertinent to what I need. I have no faith in any company actually keeping what's being put into it "safe" or "secure". It's literally the point of all this to take what people put in and use it again.

Much danger. Very dislike.

1 Like

Try feeling like that when your employer is owned by one of the biggest proponents of it. I agree completely; it's why I run everything locally or not at all.

1 Like

Found this on flathub:

I love to generate pictures with AI to make wallpapers. :slight_smile: I use DreamUp.

I like that DreamUp requires artists to opt in or else doesn't train on their art, though it seems to have a cost per generation (which makes sense for a service of course, running AI uses a lot of power and processor time).

Imaginer doesn't like me it seems. Silent failures that, when run from a terminal, show a traceback in main.py. MOST of its options seem to be server based, run from Huggingface, and require an API token to use. If someone has more luck with it, it does also support Automatic1111 and Forge, by giving it the local URL those listen to, though since starting them already gives you a web UI, I'm not sure why...

Yes, you might be better off installing Stable Diffusion or Torch.
They also refer to the same but diffusion seems more reliable - and the two forks also do not demand that you use Snap or Flatpak while they disregard the Standard packaging systems.

The tools I mentioned above with setup instructions (Automatic1111 and Forge) both run Stable Diffusion, with PyTorch as a dependency, so unless there's a front end called Torch of which I'm unaware, I already am. I gave Imaginer a try when it was mentioned, just because I hadn't seen it before, so since using it locally actually REQUIRES one of the methods I mentioned above... straight back to said methods. One day I'll take the time to get ComfyUI working again...

I took Imaginer to be the front end, actually. I had at one time installed Stable Diffusion from source, which I ran on my local machine that way. For about a day.
Then I removed it.
I also later installed diffusers (torch).

Since you asked, I needed to go to my documents and look at my notes (I developed this habit since users ask on the forum, so useful things or installs that are not .deb packages, I note the [how to] in a text and save it for reference)
What I had done, including the virtual environment:

sudo apt install python3 python3-venv python3-pip

Then installed Stable Diffusion from the source package.
HuggingFace Diffusers (torth) was

pip install diffusers[torch]

Ah, no, Imaginer was new to me. Automatic1111/Forge uses a much more complicated and extensible UI that runs in a browser once Stable Diffusion is listening on a local port.

The python installs you listed are dependencies for Automatic1111 too, though once the venv is set up, the install script installs torch via pip inside the venv. Stock Forge UI:

My first try with Stable diffusion was trying from the command line in Windows, before switching to Linux. It... was painful.

1 Like

So, I just did some playing with DreamUp and while it has a short limit on characters for the prompt - It really cannot read and understand a prompt at all. On each try, about 80% of the prompt was completely ignored. The prompt said "Two vertical towers" and it displayed three images of many towers or the interior of a building.

1 Like

It's a toy,nothing more. Granted I have been lucky with mine.

1 Like

This is true. I went into it with expectations: A very human mistake. I think I have been lucky before, but I posted very generic prompts, as well. This time, I was looking for specific details in order to snip the entire image into a series of pieces, then recombine them to make a model.

In a way, it is also a bit comforting. A.I. is not smart enough to be scary, yet.

1 Like

I can't speak for DreamUp specifically, as I haven't used it, but by default, Stable Diffusion only allows 75 tokens in a prompt, which isn't much. The tools I mentioned above extend that tremendously, and with the right extensions, allow a LOT more tailoring: prompts and exclusions by regions, grouping parts of a prompt so they apply to each other, and so on.

Yep. Even with the extensions and techniques I linked above, getting what you want still calls for a significant amount of luck, and that's before running into things like choking on memory constraints causing a model to hallucinate wildly. Admittedly, that happens with a 24 GB VRAM GPU. If a site trying to make money or promote an AI product is facing that particular constraint, they're not taking their investment seriously.

1 Like

To be honest, AI scares me these days. Every big corporation wants to introduce AI into their company workflow, no matter the cost. And unlike Star Trek, which an AI built into an android body known as Data, which uses Issac Asimov's 3 laws of robotics, the AI we have in real life, is not governed by those 3 laws at all.

As companies further develop AI to get smarter and smarter, we just might reach a tipping point, where AI begins to think for itself. (Skynet anyone?) So it scares me just how much companies like Nvidia and many other's, are pushing all this AI onto humanity, with 0-protections put in place.

In an ideal world where AI is used properly and safely, these 3 laws should be observed at all times, and these 3 laws are.........

Isaac Asimov's three laws of robotics are:

  • First Law

A robot may not harm a human or allow a human to come to harm through inaction

  • Second Law

A robot must obey orders from humans, except when those orders conflict with the First Law

  • Third Law

A robot must protect its own existence, except when doing so conflicts with the First or Second Law

[BONUS]

Some say that as AI becomes more integrated into our lives, we should continue to evolve our ethical frameworks. Some proposed additions to Asimov's laws include:

  • Fourth Law: Emphasize transparency and accountability

  • Law Zero: Ensure the welfare of humanity as a whole


1 Like

It is immensely frustrating to be bound by NDA when working for a company facing numerous layoffs, that's owned by a huge proponent of AI. Suffice it to say there's a TON I'd like to say, and that my opinion of corporations' heavy push to monetize AI is very, very dim.

2 Likes

AI grows more and more every year.

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.