MODELS
What the machinery is doing (diffusion, latent space, steps, seeds, adapters) explained in the amount of detail a producer needs and no more.
Attention
The mechanism that lets a model weigh every part of its input against every other part, and the reason long sequences cost disproportionately more.
Autoregressive Video Model
A video model that generates forward in time, each segment conditioned on what it already produced, with no single pass denoising the whole clip.
CFG Scale
A dial controlling how literally a diffusion model obeys the prompt, trading obedience against image quality.
Checkpoint
A saved snapshot of a model’s weights at a point in training, and the thing you pin when you say a pipeline is reproducible.
Context Window
The maximum amount of input a model can consider at once, measured in tokens, and a hard ceiling on how much brief it can hold.
ControlNet
A conditioning layer that constrains a diffusion model with structural input (a pose skeleton, a depth map, an edge trace) so composition survives a change of style.
Denoising Steps
How many passes a diffusion model takes to walk an image from noise to a finished frame.
Diffusion Model
The architecture behind most image and video generators: it learns to reverse a noising process, and generates by denoising random static into a frame that matches your conditioning.
Diffusion Transformer
A diffusion model whose denoising network is a transformer in place of a convolutional U-Net, which is what let video models scale.
Distillation
Training a smaller or faster model to reproduce the behaviour of a larger one, which is how most fast generation modes are built.
Embedding
A list of numbers representing a piece of text, an image or a clip, positioned so that similar things sit close together.
Fine-Tuning
Continuing a model’s training on a narrow dataset so it reliably produces one style, one subject or one house voice.
Flow Matching
A training approach that learns a straighter path from noise to image, letting a model produce good output in fewer steps.
Foundation Model
A large general-purpose model trained on broad data, intended to be adapted to specific tasks before anybody uses it.
Hallucination
Confident invention: a model producing a certification mark, a statistic, a limb or a citation that does not exist.
Inference
Running a trained model to produce an output, as distinct from training it, and the part you are billed for per use.
Latent Space
The compressed mathematical space a generative model works in, where similar concepts sit near each other.
LoRA
A small trained adapter that teaches a large model one style, character or object without retraining the model itself.
Model Licence
The terms governing what you may do with a model’s output, which differ per model, per version and sometimes per plan.
Multimodal Model
A model that takes and produces more than one kind of input (text, image, audio, video) inside a single system.
Quantisation
Storing a model’s weights at lower numeric precision so it runs on smaller hardware, at some cost to output quality.
RAG
Fetching relevant source material at run time and giving it to the model, so the answer comes from documents and not from memory.
Sampler
The algorithm that decides how a diffusion model steps from noise to image, trading speed against fidelity and consistency.
Seed
The number that fixes a model’s starting randomness, making a generation reproducible.
Temporal Coherence
Whether a generated clip holds together across time: the same face, the same coat, the same number of fingers, from first frame to last.
Text-to-Video
Generating a moving shot directly from a written description, with no source image.
Token
The unit a language model reads and writes, roughly a short word or a fragment of one. It is also the unit most models are billed by.
Tool Calling
A model requesting that a defined function be run and reading the result, which is the mechanism that turns a chat into a workflow.
Variational Autoencoder
The component that compresses an image into the compact latent space a diffusion model works in, and decodes it back to pixels afterwards.
Enough theory.
Bring us the launch and we will show you which of these terms is about to cost you a week.