Generative AI Glossary / AI Dictionary / AI Terminology

Generative AI Glossary

 Full table in the post!

Term	Tags	Description
.ckpt	Model	“Checkpoint”, a file format created by PyTorch Lightning, a PyTorch research framework. It contains a PyTorch Lightning machine learning model used (by Stable Diffusion) to generate images.
.pt	Software	A machine learning model file created using PyTorch, containing algorithms used to automatically perform a task.
.Safetensors	Model	A file format for Checkpoint models, less susceptible to embedded malicious code. See “Pickle”
ADetailer	Software, Extension	A popular Automatic1111 Extension, mostly used to enhance fine face and eye detail, but can be used to re-draw hands and full characters.
AGI	Concept	Artificial General Intelligence (AGI), the point at which AI matches or exceeds the intelligence of humans.
Algorithm	Concept, Software	A series of instructions that allow a computer to learn and analyze data, learning from it, and use that learning to interpret and accomplish future tasks on its own.
AnimateDiff	Software, Extension	Technique which involves injecting motion into txt2img (or img2img) generations. https://animatediff.github.io/
API	Software	Application Programmer Interface – a set of functions and tools which allow interaction with, or between, pieces of software.
Auto-GPT	Software, LLM
Automatic1111	Developer, SD User Interface	Creator of the popular Automatic1111 WebUI graphical user interface for SD. AUTOMATIC1111 is the de facto GUI for Stable Diffusion.
Bard	Software, LLM	Google’s Chatbot, based on their LaMDA model.
Batch		A subset of the training data used in one iteration of model training. In inference, a group of images.
Bias	Concept, LLM	In Large Language Models, errors resulting from training data; stereotypes, attributing certain characteristics to races or groups of people, etc. Bias can cause models to generate offensive and harmful content.
Bing	Software, LLM	Microsoft’s ChatGTP powered Chatbot.
CFG	Setting	Classifier Free Guidance, sometimes “Guidance Scale”. Controls how closely the image generation process follows the text prompt.
Checkpoint	Model	The product of training on millions of captioned images scraped from multiple sources on the Web. This file drives Stable Diffusion’s txt2img, img2img, txt2video
Civitai (Civitai.com)	Community Resource	Popular hosting site for all types of Generative AI resources.
Civitai Generator	Software, Tool	Free Stable Diffusion Image Generation Interface, available on Civitai.com.
Civitai Trainer	Software, Tool	LoRA Training interface, available on Civitai.com, for SDXL and 1.5 based LoRA.
CLIP	Software	An open source model created by OpenAI. Trained on millions of images and captions, it determines how well a particular caption describes an image.
Cmdr2	Developer, SD User Interface	Creator of the popular EasyDiffusion, simple one-click install graphical user interface for SD.
CodeFormer	Face/Image Restoration, Model	A facial image restoration model, for fixing blurry, grainy, or disfigured faces.
Colab	Tool	Colaboratory, a product from Google Research, allowing execution of Python code through the browser. Particularly geared towards machine learning applications. https://colab.research.google.com/
ComfyUI	SD User Interface, Software	A popular powerful modular UI for Stable Diffusion with a “workflow” type workspace. Somewhat more complex than Auto1111 WebUI https://github.com/comfyanonymous/ComfyUI
CompVis	Organization	Computer Vision & Learning research group at Ludwig Maximilian University of Munich. They host Stable Diffusion models on Hugging Face.
Conda	Application, Software	An open source package manager for many programming languages, including Python.
ControlNet	UI Extension	An Extension to Auto1111 WebUI allowing images to be manipulated in a number of ways. https://github.com/Mikubill/sd-webui-controlnet
Convergence	Concept	The point in image generation where the image no longer changes as the steps increase.
CUDA	Hardware, Software	Compute Unified Device Architecture, Nvdia’s parallel processing architecture.
DALL-E / DALL-E 2	Organization	Deep learning image models created by OpenAI, available as a commercial image generation service.
Danbooru	Community Resource	English-based image board website specializing in erotic manga fan art, NSFW.
Danbooru Tag	Community Resource	System of keywords applied to Danbooru images describing the content within. When using Checkpoint models trained on Danbooru images, keywords can be referenced in Prompts.
DDIM (Sampler)	Sampler	Denoising Diffusion Implicit Models. See Samplers.
Deep Learning	Concept	A type of Machine Learning, where neural networks attempt to mimic the behavior of the human brain to perform tasks.
Deforum	UI Extension, Community Resource	A community of AI image synthesis developers, enthusiasts, and artists, producing Generative AI tools. Most commonly known for a Stable Diffusion WebUI video extension of the same name.
Denoising/Diffusion	Concept	The process by which random noise (see Seed) is iteratively reduced into the final image.
depth2img	Concept	Infers the depth of an input image (using an existing model), and then generates new images using both the text and depth information.
Diffusion Model (DM)	Model	A generative model, used to generate data similar to the data on which they are trained.
DPM adaptive (Sampler)	Sampler	Diffusion Probabilistic Model (Adaptive). See Samplers. Ignores Step Count.
DPM Fast (Sampler)	Sampler	Diffusion Probabilistic Model (Fast). See Samplers.
DPM++ 2M (Sampler)	Sampler	Diffusion Probabilistic Model – Multi-step. Produces good quality results within 15-20 Steps.
DPM++ 2M Karras (Sampler)	Sampler	Diffusion Probabilistic Model – Multi-step. Produces good quality results within 15-20 Steps.
DPM++ 2S a Karras (Sampler)	Sampler	Diffusion Probabilistic Model – Single-step. Produces good quality results within 15-20 Steps.
DPM++ 2Sa (Sampler)	Sampler	Diffusion Probabilistic Model – Single-step. Produces good quality results within 15-20 Steps.
DPM++ SDE (Sampler)	Sampler
DPM++ SDE Karras (Sampler)	Sampler
DPM2 (Sampler)	Sampler
DPM2 a (Sampler)	Sampler
DPM2 a Karras (Sampler)	Sampler
DPM2 Karras (Sampler)	Sampler
DreamArtist	UI Extension, Software	An extension to WebUI allowing users to create trained embeddings to direct an image towards a particular style, or figure. A PyTorch implementation of the research paper DreamArtist: Towards Controllable One-Shot Text-to-Image Generation via Contrastive Prompt-Tuning, Ziyi Dong, Pengxu Wei, Liang Lin.
DreamBooth	Software, Community Resource	Developed by Google Researchers, DreamBooth is a deep learning image generation model designed to fine-tune existing models (checkpoints). Can be used to create custom models based on a set of images.
DreamStudio	Organization, SD User Interface	A commercial web-based image generation service created by Stability AI using Stable Diffusion models.
Dropout (training)	Concept	A technique to prevent overfitting by randomly ignoring some images/tokens, etc. during training.
DyLoRA C3Lier
DyLoRA LierLa
DyLoRA Lycoris
EMA	Model	Exponential Moving Average. A full EMA Checkpoint model contains extra training data which is not required for inference (generating images). Full EMA models can be used to further train a Checkpoint.
Emad	Organization, Developer	Emad Mostaque, CEO and co-founder of Stability AI, one of the companies behind Stable Diffusion.
Embedding	Model, UI Extension	Additional file inputs to help guide the diffusion model to produce images that match the prompt. Can be a graphical style, representation of a person, or object. See Textual Inversion and Aesthetic Gradient.
Emergent Behavior	Concept, LLM	Unintended abilities exhibited by an AI model.
Entropy	Concept	A measure of randomness, or disorder.
Epoch	Concept	The number of times a model training process looked through a full data set of images. E.g. The 5th Epoc of a Checkpoint model looked five times through the same data set of images.
ESRGAN	Upscaler, Model	Enhanced Super-Resolution Generative Adversarial Networks. A technique to reconstruct a higher-resolution image from a lower-resolution image. E.g. upscaling of a 720p image into 1080p. Implemented as a tool within many Stable Diffusion interfaces.
Euler (Sampler)	Sampler	Named after Leonhard Euler, a numerical procedure for solving ordinary differential equations, See Samplers.
Euler a (Sampler)	Sampler	Ancestral version of the Euler sampler. Named after Leonhard Euler, a numerical procedure for solving ordinary differential equations, See Samplers.
Finetune	Concept
float16	Setting, Model, Concept	Half-Precision floating point number.
float32	Setting, Model, Concept	Full-Precision floating point number.
Generative Adversarial Networks (GANs)	Model	A pair of AI models: one generates new data, and the other evaluates its quality.
Generative AI	Concept
GFPGAN	Face/Image Restoration, Model	Generative Facial Prior, a facial restoration model for fixing blurry, grainy, or disfigured faces.
Git (GitHub)	Application, Software	Hosting service for software development, version control, bug tracking, documentation.
GPT-3	Model, LLM	Generative Pre-trained Transformer 3, a language model, using machine learning to produce human-like text, based on an initial prompt.
GPT-4	Model, LLM	Generative Pre-trained Transformer 4, a language model, using machine learning to produce human-like text, based on an initial prompt. A huge leap in performance and reasoning capability over GPT 3/3.5.
GPU	Hardware	A Graphics Processing Unit, a type of processor designed to perform quick mathematical calculations, allowing it to render images and video for display.
Gradio	Software	A web-browser based interface framework, specifically for Machine Learning applications. Auto1111 WebUI runs in a Gradio interface.
Hallucinations (LLM)	LLM, Concept	Sometimes LLM models like ChatGPT produce information that sounds plausible but is nonsensical or entirely false. This is called a Hallucination.
Hash (Checkpoint model)	Model, Concept	An algorithm for verifying the integrity of a file, by generating an alphanumeric string unique to the file in question. Checkpoint models are hashed, and the resulting string can be used to identify that model.
Heun (Sampler)	Sampler	Named after Karl Heun, a numerical procedure for solving ordinary differential equations. See Samplers.
Hugging Face	Organization	A community/data science platform providing tools to build, train, and deploy machine learning models.
Hypernetwork (Hypernet)	Model	A method to guide a Checkpoint model towards a specific theme, object, or character based on its’ own content (no external data required).
img2img	Concept	Process to generate new images based on an input image, and txt2img prompt.
Inpainting	Concept	The practice of removing or replacing objects in an image based on a painted mask.
Kohya	Software	Can refer to Kohya-ss scripts for LoRA/finetuning (https://github.com/kohya-ss/sd-scripts) or the Windows GUI implementation of those scripts (https://github.com/bmaltais/kohya_ss)
LAION	Organization	A non-profit organization, providing data sets, tools, and models, for machine learning research.
LAION-5B	Model	A large-scale dataset for research purposes consisting of 5.85 billion CLIP-filtered image-text pairs.
Lanczos	Upscaler, Model	An interpolation method used to compute new values for sampled data. In this case, used to upscale images. Named after creator, Cornelius Lanczos.
Large Language Model (LLM)	LLM, Model	A type of Neural Network that learns to write and converse with users. Trained on billions of pieces of text, LLMs excel at producing coherent sentences and replying to prompts in the correct context. They can perform tasks such as re-writing and summarizing text, chatting about various topics, and performing research.
Latent Diffusion	Model	A type of diffusion model that contains compressed image representations instead of the actual images. This type of model allows the storage of a large amount of data that can be used by encoders to reconstruct images from textual or image inputs.
Latent Mirroring	Concept, UI Extension	Applies mirroring to the latent images mid-generation to produce anything from subtly balanced compositions to perfect reflections.
Latent Space	Concept	The information-dense space where the diffusion model’s image representation, attention, and transformation are merged and form the initial noise for the diffusion process.
LDSR	Upscaler	Latent Diffusion Super Resolution upscaling. A method to increase the dimensions/quality of images.
Lexica	Community Resource	Lexica.art, a search engine for stable diffusion art and prompts.
LlamaIndex (GPT Index)	Software, LLM	https://github.com/jerryjliu/llama_index – Allows the connection of text data to an LLM via a generated “index”.
LLM	LLM, Model	A type of Neural Network that learns to write and converse with users. Trained on billions of pieces of text, LLMs excel at producing coherent sentences and replying to prompts in the correct context. They can perform tasks such as re-writing and summarizing text, chatting about various topics, and performing research.
LMS (Sampler)	Sampler
LMS Karras (Sampler)	Sampler
LoCON
LoHa
LoKR
LoRA	Model, Concept	Low-Rank Adaptation, a method of training for SD, much like Textual Inversion. Can capture styles and subjects, producing better results in a shorter time, with smaller output files, than traditional finetuning. A LoRA is a way to get a better quality character, pose, scene, or concept than what the base model can give. A base model is the basis for generating images. It has all the trained “knowledge” (weights) about what a car looks like, what Goku looks like, what a penguin looks like, matched to the token “car”, “Goku”, “penguin”. All that was trained on billions of images and text. However it doesn’t “know” everything, or maybe doesn’t know enough about something, so sometimes you get a muddled image or not-quite lookalike. LORAs are like additional training for something specific, that let you prompt something that isn’t in the base model, or not well trained. The amazing thing is that the LoRA is mostly concepts/styles, and can “adapt” to the base model. For example, you can train a LoRA on an anime character to get their appearance and style (clothes), and then use the LoRA with a realistic base model, and get a realistic version of the character.
LoRA C3Lier
LoRA LierLa
Loss (function)	Concept	A measure of how well an AI model’s outputs match the desired outputs.
Merge (Checkpoint)	Model	A process by which Checkpoint models are combined (merged) to form new models. Depending on the merge method (see Weighted Sum, Sigmoid) and multiplier, the merged model will retain varying characteristics of its’ constituent models.
Metadata	Concept, Software	Metadata is data that describes data. In the context of Stable Diffusion, metadata is often used to describe the Prompt, Sampler settings, CFG, steps, etc. which are used to define an image, and stored in a .png header.
MidJourney	Organization, SD User Interface	A commercial web-based image generation service, similar to DALL-E, or the free, open source, Stable Diffusion.
Model	Model	Alternative term for Checkpoint
Motion Module	Software	Used by AnimateDiff to inject motion into txt2img (or img2img) generations.
Multimodal AI	Concept	AI that can process multiple types of inputs, including text, images, video or speech.
Negative Prompt	Setting, Concept	Keywords which tell a Stable Diffusion prompt what we don’t want to see, in the generated image.
Neural Network	Concept, Software	Mathematical systems that act like a human brain, with layers of artificial “neurons” helping find connections between data.
Notebook	Community Resource, Software	See Colab. A Jupyter notebook service providing access, free of charge, to computing resources including GPUs.
NovelAI (NAI)	Organization	A paid, subscription based AI-assisted story (text) writing service. Also has a txt2img model, which was leaked and is now incorporated into many Stable Diffusion models.
Olivio (Sarikas)	Community Resource	Olivio produces wonderful SD content on YouTube (https://www.youtube.com/@OlivioSarikas) – one of the best SD news YouTubers out there!
OpenAI	Organization	AI research laboratory consisting of the for-profit corporation OpenAI LP and the non-profit OpenAI Inc.
OpenPose	Model, Software	A method for extracting a “skeleton” from an image of a person, allowing poses to be transferred from one image to another. Used by ControlNet.
Outpainting	Concept	The practice of extending the outer border of an image, into blank canvas space, while maintaining the style and content of the image.
Overfitting	Concept	When an AI model learns the training data too well and performs poorly on unseen data.
Parameters (LLMs)	Concept, Software, LLM	Numerical points across a Large Language Model’s training data. Parameters dictate how proficient the model is at its tasks. E.g. a 6B (Billion) Parameter model will likely perform less well than a 13B Parameter model.
Pickle	Concept, Software	Community slang term for potentially malicious code hidden within models and embeddings. To be “pickled” is to have unwanted code execute on your machine (be hacked).
PLMS (Sampler)	Sampler	Pre-Trained Language Models. See Samplers.
Prompt	Concept	Text input to Stable Diffusion describing the particulars of the image you would like output.
Pruned/Pruning	Model	A method of optimizing a Checkpoint model to increase the speed of inference (prompt generation), file size, and VRAM cost.
Python	Application, Software	A popular, high-level, general purpose coding language.
PyTorch	Application, Software	An open source machine learning library, created by META.
Real-ESRGAN	Upscaler	An image restoration method.
Refiner	Model	Part of SDXL’s two-stage pipeline – the Refiner further enhances detail from the base model.
SadTalker	UI Extension	https://github.com/OpenTalker/SadTalker A framework for facial animation/lip synching based upon an audio input.
Samplers	Sampler	Mathematical functions providing different ways of solving differential equations. Each will produce a slightly (or significantly) different image result from the random latent noise generation.
Sampling Steps	Sampler, Concept	The number of how many steps to spend generating (diffusing) your image.
SD 1.4	Model	A latent txt2img model, the default model for SD at release. Fine-tuned on 225k steps at resolution 512×512 on laion-aesthetics v2 data set.
SD 1.5	Model	A latent txt2img model, updated version of 1.4, fine-tuned on 595k steps at resolution 512×512 on laion-aesthetics v2 data set.
SD UI	Application, Software	Colloquial term for Cmdr2’s popular graphical interface for Stable Diffusion prompting.
SD.Next	Software	See Vlad, Vladmandic Fork of Auto1111 WebUI.
SDXL 0.9	Model	Stability AI’s latest (March 2023) Stable Diffusion Model. Will become SDXL 1.0 and be released ~July 2023.
Seed	Concept	A pseudo-random number used to initialize the generation of random noise, from which the final image is built. Seeds can be saved and used along with other settings to recreate a particular image.
Shoggoth Tongue	Concept, LLM	A humorous allusion to the language of the fictional monsters in the Cthulhu Mythos, “Shoggoth Tongue” is the name given to advanced ChatGPT commands which are particularly arcane and difficult to understand, but allow ChatGPT to perform advanced actions outside of the intended operation of the system.
Sigmoid (Interpolation Method)	Model, Concept	A method for merging Checkpoint Models based on a Sigmoid function – a mathematical function producing an “S” shaped curve.
Stability AI	Organization	AI technology company co-founded by Emad Mostaque. One of the companies behind Stable Diffusion.
Stable Diffusion (SD)	Application, Software	A deep learning, text-to-image model released in 2022. It is primarily used to generate detailed images based on provided text descriptions.
SwinIR	Face/Image Restoration, Model	An image restoration transform, aiming to restore high quality images from low quality images.
Tensor	Software	A container, in which multi-dimensional data can be stored.
Tensor Core	Hardware	Processing unit technology developed by Nvidia, designed to carry out matrix multiplication, an arithmetic operation.
Textual Inversion	Model, Concept, UI Extension	A technique for capturing concepts from a small number of sample images in a way that can influence txt2img results towards a particular face, or object.
Token	Concept	A token is roughly a word, a punctuation, or a Unicode character in a prompt.
Tokenizer	Concept, Model	The process/model through which text prompts are turned into tokens, for processing.
Torch 2.0	Software	The latest (March 2023) PyTorch release.
Training	Concept	The process of teaching an AI model by feeding it data and adjusting its parameters.
Training Data	Model	A set of many images used to “train” a Stable Diffusion model, or embedding.
Training Data	Concept, LLM, Model	The data sets uses to help AI models learn; can be text, images, code, or other data, depending on the type of model to be trained.
Turing Test	Concept	Named after mathematician Alan Turing, a test of a machine’s ability to behave like a human. The machine passes if a human can’t distinguish the machine’s response from another human.
txt2img	Concept, Model	Model/method of image generation via entry of text input.
txt2video	Concept, Model	Model/method of video generation via entry of text input.
Underfitting		When an AI model cannot capture the underlying pattern of the data due to incomplete training.
UniPC (Sampler)	Sampler	A recently released (3/2023) sampler based upon https://huggingface.co/docs/diffusers/api/schedulers/unipc
Upscale	Upscaler, Concept	The process of converting low resolution media (images or video) into higher resolution media.
VAE	Model	The Variational AutoEncoder converts the image between the pixel and the latent spaces. It basically encodes and decodes data into visuals. Variational Autoencoder. A .vae.pt file which accompanies a Checkpoint model and provides additional detail improvements. Not all Checkpoints have an associated vae file, and some vae files are generic and can be used to improve any Checkpoint model.
Vector (Prompt Word)	Concept	An attempt to mathematically represent the meaning of a word, for processing in Stable Diffusion.
Venv	Software	A Python “Virtual Environment” which allows multiple instances of python packages to run, independently, on the same PC.
Vicuna	LLM, Software, Model	https://vicuna.lmsys.org/ An Open-Source Chatbot model founded by students and faculty from UC Berkeley in collaboration with UCSD and CMU.
Vladmandic	Software, SD User Interface	A popular “Fork” of Auto1111 WebUI, with its own feature-set. https://github.com/vladmandic/automatic
VRAM	Hardware	Video random access memory. Dedicated Graphics Card (GPU) memory used to store pixels, and other graphical processing data, for display.
Waifu Diffusion	Model	A popular text-to-image model, trained on high quality anime images, which produces anime style image outputs. Originally produced for SD 1.4, now has an SDXL version.
WebUI	Application, Software, SD User Interface	Colloquial term for Automatic1111’s WebUI – a popular graphical interface for Stable Diffusion prompting.
Weighted Sum (Interpolation Method)	Concept	A method of Checkpoint merging using the formula Result = ( A * (1 – M) ) + ( B * M ) .
Weights	Model	Alternative term for Checkpoint
Wildcards	Concept	Text files containing terms (clothing types, cities, weather conditions, etc.) which can be automatically input into image prompts, for a huge variety of dynamic images.
xformers	UI Extension, Software	Optional library to speed up image generation. Superseded somewhat by new options implemented by Torch 2.0
yaml	Software, UI Extension, Model	A human-readable data-serialization programming language commonly used for configuration files. Yaml files accompany Checkpoint models, and provide Stable Diffusion with additional information about the Checkpoint.