Glossary

Age Verification

The process a platform uses to confirm a user meets a minimum age before granting access to adult content, ranging from a simple checkbox to document or biometric checks.

AI Companion

An AI application built around an ongoing, personalized relationship with a persistent virtual character: the umbrella category that AI girlfriend, AI boyfriend, and general companion chatbot apps all fall under.

AI Roleplay

A style of interacting with an AI (writing collaborative, in-character fiction turn by turn) rather than a product; it can happen inside a dedicated companion app or in a general-purpose chatbot.

API

A defined interface that lets one piece of software send requests to and get results back from another system, such as a hosted AI model, usually over the internet.

Attention Mechanism

A learned operation that weights which other tokens or features matter when updating the current representation.

Base Model

The underlying pretrained model from which chat variants, fine-tunes, adapters, and merges are derived.

Character Card

A portable bundle of character instructions and metadata that character-chat frontends use to recreate a persona, scenario, greeting, and example dialogue.

Character Chatbot

The general product type behind AI companions: a chat interface built around a persistent character persona, which may or may not be romantic: from an original companion to a fictional or public-figure-styled bot.

Character Consistency

The goal of keeping a character’s face and appearance stable across many separate generations or video frames: a problem addressed by combining several techniques, not any single one.

Character Memory

A companion app's user-facing feature for tracking details about you and your relationship with a specific character, built on top of the context window and/or longer-term storage.

Chat Template

A model-specific formatting recipe that serializes roles and messages into the exact token sequence expected by a chat-tuned language model.

Checkpoint

A single saved file of a model's weights that generation software loads and runs: the concrete file, not the abstract concept of "weights" or a small add-on like a LoRA.

CLIP

The model that converts text (and images) into the embeddings a diffusion model uses to understand and follow a prompt.

Conditioning

Information supplied to steer a generative model, such as text, a source image, mask, pose, depth map, reference embedding, or motion signal.

Content Credentials (C2PA)

Cryptographically signed metadata, defined by the C2PA specification, recording a media file's creation and edit history: provenance information, not an AI detector.

Content Moderation

A platform's overall process, policy, and staffing for reviewing, restricting, or removing content: the broad practice that tools like an NSFW filter are one part of.

Context Window

The fixed amount of text, measured in tokens, a model can process in a single request: a hard technical ceiling on how much of a conversation it can see at once.

ControlNet

Adds an extra structural input (an edge map, depth map, or pose skeleton) to a diffusion model so a generation’s composition can be controlled independently of the text prompt.

Credits

A prepaid, spendable unit that hosted AI platforms use to meter usage instead of charging per generation in real currency.

Cross-Attention

An attention operation in which one representation queries information from another, commonly connecting prompt embeddings to an image denoiser.

Deepfake

AI-manipulated media that depicts a real, identifiable person doing or saying something they didn't: narrower than synthetic media, which also covers entirely invented characters.

Denoising Strength

An image-to-image control that sets how much noise is added to the starting latent and therefore how far the result may depart from the source image.

Diffusion Model

The general class of generative model that creates images, video, or audio by starting from random noise and iteratively removing it, step by step, until a coherent output remains.

Diffusion Transformer (DiT)

A diffusion or flow-model denoiser built from transformer blocks operating on image latents or patches instead of a conventional U-Net backbone.

DreamBooth

A fine-tuning method that bakes a specific subject into a model's actual weights through further training, typically producing a new full or merged checkpoint: heavier and more permanent than a LoRA.

Embedding

A numerical vector that represents the meaning of a piece of text, image, or other input, letting a model compare or condition on things by similarity rather than exact matches.

Face Swap

A face-synthesis or editing process that transfers a source identity into a target image or video while attempting to retain pose, expression, lighting, and scene context.

Frame Interpolation

Fills in the frames between two existing frames, used to smooth motion or raise a clip’s frame rate rather than generate a new sequence from scratch.

Generative AI

The broad category of AI systems trained to create new text, images, video, or audio, rather than to classify or predict on data that already exists.

GGUF

A single-file binary format from the llama.cpp/GGML project that packages a local model's weights and metadata together for fast loading: the common container for quantized local models, but not the same thing as quantization itself.

GPU Inference

Running a trained model's forward computations on a graphics processor to exploit highly parallel matrix operations and fast device memory.

Group Chat

A multi-character conversation mode that coordinates two or more AI personas in one shared chat context.

Guidance Scale / CFG

The setting that controls how strongly generation follows the prompt, where higher values mean stricter prompt adherence but not necessarily better images.

Hallucination

A known LLM failure mode in which a model confidently states something false, fabricated, or inconsistent with earlier facts, not a memory setting or a censorship issue.

Hires Fix

A two-pass Stable Diffusion workflow that generates at a manageable resolution, upscales, then denoises the enlarged image to add coherent detail.

Hosted Generation

Using an AI model that runs on a company's remote servers and is accessed through a website, app, or API, rather than on your own hardware.

Image-to-Image

Regenerating an entire image guided by both a starting image and a prompt, at an adjustable strength that trades fidelity to the original for creative freedom.

Image-to-Video

Turning a single still image into a short animated or video clip using a generative model.

Inference

The step where an already-trained model is run to produce an output, as distinct from training, the earlier process of teaching the model its parameters.

Inference Steps

The number of denoising iterations a generation runs, where more isn't automatically better past a model-dependent point.

Inpainting

Editing only a masked region of an existing image while the rest of the picture stays untouched.

IP-Adapter

Conditions a generation on a reference image’s visual content or style through a lightweight image-encoder and dedicated cross-attention layers, without retraining the base model.

Keyframe

An anchor frame (typically the first or last frame of a clip) that’s fully generated or specified, pinning down the video at a fixed point in time.

KV Cache

A language-model inference cache that reuses attention keys and values from earlier tokens instead of recomputing the entire prefix for every new token.

Large Language Model (LLM)

The underlying neural-network technology (trained to predict the next chunk of text) that powers character chatbots, AI companions, and most modern conversational AI.

Latent Diffusion

The specific, now-dominant diffusion technique that runs the noise-removal process inside a compressed latent space instead of on raw pixels, making generation far cheaper.

Latent Space

The compressed, lower-dimensional numerical space a model learns to represent complex data in, where similar items end up positioned near each other.

Likeness Consent

Clear permission from an identifiable person to capture, transform, generate, or distribute their likeness for a defined purpose and context.

Lip Sync

Matches a character’s mouth movements in a video to a specific audio track, whether generated speech, a cloned voice, or a separately recorded voiceover.

Local Generation

Running an AI model for chat, images, or video entirely on your own device's hardware, instead of sending requests to a company's remote servers.

Long-Term Memory

A mechanism for storing information outside the model's context window so it survives across sessions, rather than disappearing once a conversation exceeds the model's limit.

LoRA

A small (typically a few MB to a few hundred MB) add-on file that adjusts a frozen base checkpoint toward a character, style, or concept, not a standalone model on its own.

Model Card

A document published with a trained model to report its intended uses, technical context, evaluation, limitations, and other release metadata.

Model Distillation

A training process in which a student model learns to reproduce selected behavior from a teacher model or more expensive generation process.

Model Fine-Tuning

The general activity of further-training an already-trained model on new data to specialize its behavior: the umbrella term covering specific methods like DreamBooth, LoRA, and Textual Inversion.

Model Marketplace

A discovery and distribution platform for model repositories, checkpoints, adapters, and metadata, whether files are shared freely or under commercial access terms.

Model Merge

A method of combining parameters or learned adapters from compatible models to produce one model with a blended set of behaviors.

Model Offloading

A memory-management technique that moves model components or caches between GPU memory, system RAM, and sometimes storage so oversized workloads can run.

Model Weights

The numerical parameters a neural network learned during training: the underlying data that gets packaged into a checkpoint file for a tool to load and run.

Motion Control

The inputs (camera paths, motion-strength sliders, drawn trajectories) used to directly steer how a generated video moves, instead of leaving motion entirely up to the model.

Multimodal AI

An AI system that processes or produces more than one data modality (such as text, images, audio, or video) within a coordinated model or workflow.

Negative Prompt

A separate field describing what to exclude from a generation, steering the model away from unwanted content or quality issues.

NSFW Filter

An automated classifier or ruleset that flags or blocks sexually explicit content: one specific technical component used inside a platform's broader content moderation system.

Numerical Precision

The number format and bit width used to store or compute model values, such as FP32, FP16, or BF16, with direct effects on memory, speed, and numerical range.

Open-Weight Model

A model whose trained weights are published for anyone to download and run on their own hardware, as opposed to a model only reachable through a company's hosted API.

Outpainting

Extending an image beyond its original borders, generating new surrounding content rather than editing what's already there.

Persona

The defined identity (name, personality traits, backstory, speech style) assigned to a chatbot so it consistently behaves as a specific character rather than a generic assistant.

Prompt

The text, image, or other input that tells a generative model what to produce or how to respond.

Prompt Weighting

Syntax or controls that increase or decrease the relative influence of selected prompt fragments during conditioned generation.

Quantization

Reducing the numerical precision a model's weights are stored in (e.g., 16-bit down to 4-bit) to shrink file size and compute needs: a general technique, distinct from any one file format.

Rectified Flow

A generative-model formulation that learns a velocity field for transporting noise toward data, ideally along paths that can be integrated efficiently during inference.

Reference Image

Any image supplied as an input to steer a generation’s subject, style, or composition: the general concept behind IP-Adapter, image-to-image, and image-to-video.

Retrieval-Augmented Generation (RAG)

A specific technique for feeding a model outside information at answer time (retrieving relevant stored text and inserting it into the prompt) commonly used to implement long-term memory and reduce hallucination.

SafeTensors

A file format for storing model weights safely as an alternative to Python's pickle-based .ckpt/.pt files, independent of whether the weights inside are a base checkpoint, a fine-tune, or a LoRA.

Sampler

The step-by-step algorithm (Euler, DPM++, etc.) that performs the actual denoising at each generation step.

Scheduler

The setting that controls how noise levels are spaced across a generation's steps, which the sampler algorithm then follows.

Seed

The number that initializes a generation's starting noise, which a user saves to reproduce the exact same result again.

Synthetic Media

The broad umbrella term for any media wholly or partly generated or altered by AI, including entirely original characters that have no real-person basis at all.

System Prompt

The hidden instructions sent to a language model before every conversation that implement a persona's behavior, rules, and boundaries under the hood.

Temperature

A generation setting that controls how random or predictable a model's word choices are: low values give consistent, repetitive output, high values give more varied, sometimes erratic output.

Temporal Consistency

How stable a subject’s appearance and geometry stay from one video frame to the next, as opposed to flickering, warping, or drifting over time.

Text Encoder

A model component that turns tokenized prompt text into numerical representations a generator can condition on.

Text-to-Image

Generating a complete image directly from a written prompt, with no starting image involved.

Text-to-Speech (TTS)

A speech-synthesis system that converts written dialogue into audio, often adding selectable voices to companion and character chats.

Text-to-Video

Generates a video clip directly from a text prompt alone, with no starting image or existing footage involved.

Textual Inversion

A technique that teaches a model a new word (a learned text embedding) tied to a concept from a handful of images, without touching any of the model's weights at all: the lightest-weight of the customization methods.

Token

The basic chunk of text (roughly a word or word-piece) that a language model reads, generates, and is billed by; the unit that fills up a context window.

Tokenizer

A model-specific text-processing component that converts raw text into token IDs and converts generated IDs back into readable text.

Training Data

The examples used to adjust a model's learned parameters during pretraining, fine-tuning, or another training stage.

Transformer

A neural-network architecture that processes relationships among token-like inputs with attention, forming the backbone of most modern LLMs and many newer image generators.

U-Net

An encoder-decoder neural network with skip connections, widely used as the denoising backbone in Stable Diffusion 1.x and SDXL.

Uncensored Model

A language model, usually open-weight, that has had its built-in refusal and safety training removed or never applied, so it will engage with content mainstream hosted models decline.

Upscaling / Super-Resolution

Increasing an image's resolution using a model that generates plausible new detail, rather than simply stretching existing pixels.

VAE

The component that encodes images into the compressed latent space a diffusion model uses, and decodes the result back into pixels.

Video-to-Video

Transforms an existing video clip (restyling it, altering a subject, or replacing elements) while generally preserving its original motion and timing.

VRAM

The memory attached to a graphics processor that holds model weights, activations, caches, and working data during GPU inference.

Watermarking

Embedding a visible or invisible marker directly into generated media's pixels or audio to signal AI origin: a different, less structured mechanism than C2PA's signed metadata.