Basic AI Terminology for Beginners

So I've seen a good amount of people who are confused by a lot of the AI terminology out there, and I get it - it's easy to be overwhelmed by the sheer amount of new information being thrown out at you when you're just getting started in the realm of AI.

This blog will be updated periodically as I think of more stuff to add.

---

LLM

LLM stands for Large Language Model.

You've probably heard of ChatGPT, right? ChatGPT is a large language model, along with Gemini, Claude and so many others.

Open Source / Closed Source

Closed source means the source files aren't available to the public, so we can't adjust or finetune the model. ChatGPT is an example of a closed source model.

However! OpenAI has stated that they will be releasing an open source model this summer!

Exciting, right?

But what is open source?

Open source means it's available to the public, for free. We can download it, finetune it, do whatever we want (check Licenses first of course)! IT'S FREE!

Looking for open source models? Check out Huggingface, which is the best online hub for open source AI stuff. There's enough to keep you busy here for ages!

Finetune

A finetune is a base model that someone has re-trained on their own dataset (whether images, text, etc). Finetunes influence the model to generate differently based on the new data. For example, finetuning an image base model, like SDXL, with anime art is how you would get a model that is better at generating anime style images.

Base Model

Base models are just that, a base model in its unedited form, original. Base models are things like SDXL, SD1.5, Pony, Illustrious, Flux and Noobai (for image), and Gemma or Llama models (for text). Base models are very expensive to train and require huge amounts of data! For that reason, base models are usually made by big companies.

Hallucination

What does it mean when a model hallucinates? This term is generally used in reference to LLMs. A hallucination is exactly what it sounds like: the LLM is 'hallucinating' something that is untrue but presenting it as factual data.

A common one I've personally seen is an LLM hallucinating fake URLs in its output. This is because there are probably a lot of instances of people writing URLs in their posts and those posts getting into datasets en masse.

You can see some LLMs hallucinating things like Reddit-esque language, fake URLs or OOC chats that seem completely bizarre and out of nowhere.

Weight (image model) When making images, some people say 'raise the weight of a tag' or something similar. What does that mean? Simply put, weight is just a way to give a particular tag "more weight" or more importance within a prompt (the opposite can also be true.)

If you're making a picture with SDXL, you might have brown dog in your prompt.

Maybe the model keeps making the dog's fur black.

So you can put (brown:1.5) dog in your prompt to increase the likelihood of the dog having brown fur. Experiment with weights and see how it works for yourself!

Positive and Negative Prompt

A positive prompt is what you do want in your final image.

A negative prompt is what you don't want in your image.

So if you want a picture of a dog with brown fur, that goes in the positive prompt.

If you want to further reinforce that the dog does not have black fur, you could put black fur in the negative prompt.

Lots of people like to do this with quality tags as well, such as: masterpiece, best quality in the positive prompt, and worst quality, lowres in the negative prompt.

These are just a few examples. The variations you can do with prompting are endless.

Model Weights

You might also hear 'weights' in reference to a model. You can think of weights like a model's core training - without these, the model would no longer be the same.

If you want more in-depth reading, check out this wikipedia page about neutral networks.

LoRA

LoRA stands for Low Rank Adaptation. I like to think of these as little "expansion packs" that you can attach to your main model when generating images. LoRAs are essentially small models that are trained on a particular element such as a character, a style or a concept (like a pose). You can train these on your own computer if you have enough VRAM, or use a service like civitAI to make it.

You can also make LoRAs for text and video models, though I don't have any experience with that (yet)!

Checkpoint

Generally, image generation models are called checkpoints.

Inpainting

Inpainting is a feature which allows you to select (via a sort of paintbrush tool) an area of a picture to change or regenerate. The closest idea I can give you if you have no prior knowledge is Photoshop's Content Aware Tool, but much more advanced.

CFG

When generating images, you may see a parameter called CFG. What does that mean?

CFG stands for classifier-free guidance scale.

Basically, the CFG determines how closely the image generation process will follow your prompt. A higher value means more strict adherence, while a lower value offers the model more 'creative freedom'. Generally, these values are kept between 3 and 7, but experimentation can yield interesting (or disastrous) results.

That's all for now! When I think of more terms to add, they'll appear here.

Hope it helps someone. :)