Glossary: The vocabulary of generative models

Introduction

Most of these terms name one part of the same sequence. Your prompt is cut into pieces by a tokenizer and read by a text encoder, which turns it into an embedding. That embedding is the conditioning the generator is steered by, through attention, while it runs denoising: a loop that starts from noise and removes a little of it at each inference step, with a sampler deciding how much and guidance deciding how hard to push. All of that happens on a latent, not on pixels, and the VAE decoder is what turns the finished latent into the picture you receive.

The rest of the vocabulary splits in two. Some terms describe what a model is: its architecture, the checkpoint or fine-tune you address, and whether it was built on diffusion, flow matching or an autoregressive design. The others describe what you hand it: a seed, a number of steps, a mask, a reference image, or an attachment like a LoRA or a ControlNet.

Each entry below says what a term is and nothing else. How to use the parameter it belongs to lives on its own page, and the exact values it accepts live in the API reference.

Architecture

The network design a model is built on: which network does the denoising (U-Net or DiT), which text encoder reads the prompt, which VAE moves between pixels and latents, and what objective the whole thing was trained on. Everything else follows from it, including which parameters exist, which sizes work, and what a LoRA can attach to. Every checkpoint trained from one architecture shares all of that.

Aspect ratio

The proportion between width and height. A model learned what a plausible image looks like at the shapes it was trained on, so an unfamiliar ratio costs coherence before it costs anything else: handed a canvas it does not recognize, a model fills the extra room with more of what it knows. It follows from the width and height you request.

Attention

The operation that lets a model weigh every part of its input against every other part when deciding what to produce. In a generation model it is what connects a word in your prompt to a region of the image. Cross-attention relates the prompt to the image, self-attention relates parts of the image to each other, and in video models temporal attention relates one frame to the others.

Autoregressive

A model that produces its output one piece at a time, each piece conditioned on the ones already produced. Language models work this way, and so do some image and video models. It is the alternative to diffusion, which shapes the whole output at once instead of extending it.

CFG scale

The multiplier on guidance. Each step runs the model with your conditioning and without it, and this value decides how far to push the result along the difference between the two answers. Too low and the model produces whatever it finds likely, too high and it amplifies an estimate until contrast and detail flatten.

Checkpoint

A file holding a model's trained weights. On this platform a checkpoint is the thing you address with model, and community checkpoints are fine-tunes of a base architecture, which is why one accepts the same parameters as the family it came from.

Classifier-free guidance

The technique behind guidance. A model is trained to work both with conditioning and without it, so at generation time it can be run twice per step and the difference between the two answers amplified. No separate classifier is involved, which is where the name comes from.

CLIP

Contrastive Language-Image Pre-training. Two networks trained together on images and their captions, one reading pictures and one reading text, until both land matching pairs at the same place in a shared space. Generation models reuse the text half as their text encoder, which is why a short caption-shaped phrase reaches a CLIP-based model better than a long sentence does.

CLIP skip

Reads the prompt embedding from a layer before the last one of the CLIP encoder. The later layers are the most specialized, so stopping early hands the generator a looser, less literal reading of the same words. It exists only on architectures whose reader is CLIP.

Conditioning

Anything you give a model to steer what it produces, as opposed to the randomness it starts from. A prompt is conditioning, so is a reference image, a mask, a depth map or a first frame. The whole catalog of techniques is largely a catalog of different things to condition on.

ControlNet

A second network trained alongside a frozen base model to take a structural input, such as an edge map, a depth map or a pose, and feed it into the denoising network at every step. The result follows the shape you supplied while the prompt decides everything else, which makes a ControlNet a structural lock rather than a style hint.

Cross-attention

The attention that relates your prompt to the image. At every step each region of the latent weighs the prompt token embeddings and decides what it should become. It is the junction between the two spaces, and it is why one word can change the whole frame instead of one corner of it.

Denoising

The process that turns noise into an output. The model is handed something random and asked what the signal underneath would look like, then that answer is partly applied and the question is asked again. Each repetition is one inference step, and the schedule that decides how much to remove each time is the sampler.

Diffusion

A family of models trained by adding noise to real data until it is unrecognizable, then learning to reverse that corruption. Generating is running the reverse direction from pure noise. Most image models you will meet are diffusion models, and so are many video models.

Distillation

Training a small or fast model to reproduce the behavior of a slower one. A distilled model reaches a usable result in a handful of inference steps instead of dozens. Guidance distillation is the specific case where the effect of guidance is baked into the weights, which is why those models accept a CFGScale that does comparatively little.

DiT

Diffusion Transformer. A diffusion model whose denoising network is a transformer rather than a U-Net. It is what most recent architectures use, because transformers scale with data and compute in ways the convolutional design did not.

Embedding

A list of numbers that represents something in a way a model can work with. Your prompt becomes an embedding before the generator ever sees it. The word also names a specific technique, a small trained file that teaches a model one new concept through a token you can then use in a prompt.

Fine-tune

A model trained further on a narrower set of images than the one it came from. It keeps the architecture it started on, which is why a fine-tune accepts the same parameters as its base. Most community checkpoints are fine-tunes.

Flow matching

A training approach that teaches a model the direct path from noise to data rather than a sequence of denoising steps. It reaches a result in fewer steps than classical diffusion and is what several recent architectures use.

Guidance

How hard a model is pushed to follow your conditioning rather than produce whatever it finds most likely. Classically this is done by running the model twice per step, once with your prompt and once without, and amplifying the difference. CFGScale is the dial on that amplification.

Image to image

Starting generation from an existing picture rather than from pure noise. The picture is encoded, noised, and rebuilt, and strength sets how much of it is destroyed before the rebuild.

Image to video

Generating a clip that begins on a picture you supply. The image is encoded and written into one slice of the latent volume, that slice is held fixed, and the rest is denoised around it, which is why the subject and the light carry forward without you describing them.

Inference step

One pass of denoising. More steps mean more, smaller corrections, which usually means more detail up to the point where the output stops changing. A distilled model needs far fewer of them.

Inpainting

Regenerating part of an image while the rest is held fixed. Noise is added everywhere, but at every step the unmasked region is re-imposed from the original, so only what the mask marks is actually rebuilt. That is what separates an edit from a new generation.

IP-Adapter

An adapter that lets a model be conditioned on a picture as well as on text. The reference is read by an image encoder and injected through the same cross-attention the prompt uses, so it carries a style, a palette or a subject into a new composition without fixing the layout.

Keyframe

A frame whose content is fixed before the rest of a clip is generated. A first frame and a last frame are keyframes, and a few models take one at an arbitrary index, which turns the frame array into a short timeline rather than a pair of ends.

Latent

A compressed representation of an image, which is what a model shapes instead of pixels. A 1024 x 1024 image is roughly 128 x 128 as a latent, so there is far less of it to work on, and it stays laid out like the picture: a cell of the latent corresponds to a patch of the image.

Latent space

The space those compressed representations live in. It keeps the spatial layout of the picture rather than organizing images by meaning, which is what a text encoder does. Generation happens entirely here, and the picture only exists once the VAE decodes the result.

Latent volume

The video equivalent of a latent: frames, height and width together as one tensor. A video model denoises the whole volume at once rather than making images one after another, which is why the subject and the lighting carry across a clip on their own. A supplied first frame fixes one slice of that volume and the rest is generated around it.

LoRA

Low-Rank Adaptation. A small file of weight adjustments added to a base model at generation time, trained on a narrow set of images to push it toward one style, subject or concept. It is a fine-tune you can attach and detach, which is why several can be stacked and why one only fits the architecture it was trained against.

Mask

A black and white image marking which part of a picture may change: white is free, black is held. Its edges matter as much as its shape, because a hard boundary leaves a visible seam where the rebuilt region meets the original. It is what inpainting works from.

Negative prompt

Content given to the unconditioned half of guidance, so the model is pushed away from it rather than toward it. It needs that second pass to exist, which is why it does nothing on a guidance-distilled model.

Noise

Random values with no structure. Diffusion models start from pure noise and remove it until an output remains, and the seed is what decides which noise you start from.

Noise schedule

The list of noise levels a generation passes through, from pure noise down to none. It decides how large each hop is, and most schedules take large hops early and small ones late, because the trajectory bends hardest near pure noise. The sampler is what walks it.

Outpainting

Generating beyond the edges of an existing image. The original sits inside a larger canvas and the new area is the region to fill, so the model continues the picture rather than replacing it. What it continues is only what the original frame showed.

Preprocessor

A model that turns an ordinary image into the structural map a ControlNet expects: edges, a depth map, a pose skeleton, a segmentation. It runs as its own call before the generation, and which one you run decides what the ControlNet is able to lock.

Prompt

The text a model is conditioned on. It is cut up by a tokenizer, turned into an embedding by a text encoder, and read through cross-attention at every step, so how you phrase it matters in ways that depend on which encoder the architecture uses.

Quantization

Storing a model weights at lower numeric precision, such as fp8 in place of fp16, to cut memory and speed up inference. It costs some fidelity, and it is why a component can misbehave at one precision and not another, which is the usual reason a VAE produces black patches until it is swapped.

Refiner

A second model run over the last part of the denoising to add detail the base model left coarse. It receives the partly denoised latent rather than a finished image, so the handover happens inside the generation and not after it.

Sampler

The algorithm that decides how much noise to remove at each inference step and how to use the model's prediction. Different samplers reach different results from the same starting point, and some add fresh randomness each step, which is why a fixed seed does not always reproduce exactly. Also called a scheduler.

Scheduler

Another name for the sampler, and the one the parameter uses. Strictly the schedule is the list of noise levels and the sampler is the algorithm that walks it, but in practice the two words are interchangeable.

Seed

The number that picks which noise a generation starts from. The same seed with the same request reaches the same result, which is what makes a generation reproducible, and changing only the seed is how you get variations of one idea rather than a different idea.

Steps

How many inference steps a generation runs. Each one is a small correction, so more of them follow the trajectory more closely, up to the count where the remaining error is smaller than the decoder can express and the output stops changing.

Strength

How much noise is added to your image before a model rebuilds it, which decides how much of the original survives. Low values keep it, high values leave little more than the composition.

T5

A text encoder from a family of general language models, used by several recent architectures instead of or alongside CLIP. It reads long, grammatical prompts better than CLIP does, which is why prompting style differs between architectures.

Temporal layer

The part of a video model that relates frames to one another, through attention or convolution across time. It is what a video architecture has and an image architecture does not, and it is what keeps a subject consistent from one frame to the next.

Text encoder

The model that turns your prompt into an embedding the generator can be conditioned on. Which encoder an architecture uses decides how your prompt is read: CLIP reads short tag-like phrases well, while the T5 family handles long natural sentences.

Text to image

Generating from a prompt alone, starting from pure noise picked by the seed. Nothing constrains where the result lands except the conditioning, which is why it is the mode with the most variance between runs.

Textual inversion

A training technique that teaches a text encoder one new token standing for a concept, without changing the generator at all. The result is a small embedding file, and because the token lives in one encoder vocabulary, it only works on the architecture it was trained against.

Timestep

Where a latent sits along the noise schedule, meaning how much noise it still holds. The model is told the timestep with every prediction, because the same latent means different things depending on how far along it is, and strength works by choosing which timestep to start from.

Tokenizer

The step that cuts your prompt into the units the text encoder actually consumes. Those units are not words: a long or unusual word becomes several tokens, which is part of why a model handles some terms better than others.

U-Net

The denoising network most early diffusion models used, shaped so that the input is progressively compressed and then expanded back, with connections across. It has largely been replaced by DiT in recent architectures.

Upscaling

Producing a larger version of an image after it has been generated. It is a separate operation from generating at a larger size, which asks the model for a latent at a shape it may never have been trained on.

VAE

Variational Autoencoder. The pair of networks that move between pixels and latents: an encoder that compresses an image into the space the model works in, and a decoder that turns the result back into a picture. The decoder is the last step of every generation, which is why swapping it changes color and fine detail.