The vocabulary the columns are written in

Every page here explains one word that turns up in a repository or on a model card, says what it fixes and what it leaves moving, and stops. No vendor value appears on these pages, so nothing here can go stale when a repository is edited. As of 2026-09-22.

Fig. 1 Filing each note under one column is what stops a single definition from being quietly wrong in two places.

Model cards are written by the people who trained the model, for people who train models. The words are precise and the precision is invisible to a reader who arrives wanting to know whether something will run. Peak memory, offloading, precision, frame count: each has an exact meaning, and each is routinely read as a rougher claim than it is.

These notes exist so the entries can stay short. A cell in the register quotes a vendor and links here rather than explaining the unit inside the quotation, which keeps the quoted value verbatim and keeps the explanation in one place where it can be corrected once.

Each note is filed under the column it serves, because the same word can mean different things in different columns. A count in the size column is a property of a file; a count in the length column is a property of an output; and the two are confused often enough to be worth separating on purpose.

1Terms behind the licence column

Permissive licence — A licence whose text is fixed and widely known, so a reader who has cleared it once has cleared every release that carries the same name.
Vendor-written licence — A licence named after the company that wrote it, published because the company wanted to say something a standard document does not say.
Acceptable-use policy — A separate file that restricts what a model may be used for, which can sit next to a standard permissive licence without changing its name.
Redistribution — Passing weights on in any form, including inside a container image or a product, which is where licence terms start to differ from each other.
Derivative weights — Any file produced by fine-tuning, merging or quantising a released model, which is a new artefact governed by the terms of the one it came from.
Output rights — What may be done with the video a model produces, which is the question most readers arrive with and the one a weights licence often does not answer.
Licence tag — The short string in a model card's licence field, which is what this column quotes and which is not the same thing as the licence itself.

2Terms behind the size column

Parameter count — The number of learned values in a released model, which pins down the file on disk and predicts the memory a run consumes surprisingly badly.
Checkpoint — A saved set of weights at one point in training, which is the thing a reader actually downloads and the thing a figure should be attached to.
Model variant — A sibling of the main model published in the same release, differing in size, precision or training, and carrying its own figures.
Distillation — Training a smaller network to imitate a larger one, which lowers the memory and the runtime and produces a model with its own behaviour.
Quantisation — Storing weights in a smaller numeric format, which lowers the memory a run needs without changing the parameter count at all.
Autoencoder — The component that compresses frames into a latent space and back, often published as a separate parameter count because it is a separate trained model.
Diffusion transformer — The component that does the generating in most current video models, labelled DiT in a parameter breakdown, which names the design while stating the size.

3Terms behind the hardware column

Peak memory — The high-water mark a run reached on the card it was measured on, which tells a reader that machine finished and not that a smaller one would.
Memory floor — The smallest card a vendor says will complete a described job, which is only checkable when the job and the card generation are described too.
Offloading — Moving parts of a model out to system memory between steps, which lowers the card requirement, lengthens the run, and depends on the host machine.
Numeric precision — How many bytes each weight occupies, which sets the weights footprint and decides whether a given card can run the model at all.
Model sharding — Splitting one model across several cards, which lowers the memory each card needs without dividing the total requirement by the card count.
GPU architecture — A generation of chip rather than a product name, which covers cards that did not exist when the model card was written, and names the real limit.
Batch size — How many clips a run generates at once, which multiplies the activation memory and is usually left unstated beside a published figure.
Cold start — Loading tens of gigabytes of weights from storage into a card, which happens on every run and appears in no memory figure.
Throughput — How long a run takes on a stated card at a stated configuration, which is a planning reference rather than a result that can be ranked.
Storage footprint — The disk the weights occupy and the speed that disk has to sustain, neither of which appears in a parameter count or a memory figure.

4Terms behind the resolution column

Resolution tier — A label such as 540P or 720P, which means different pixel counts to different vendors and is only a value once the dimensions are written out.
Aspect ratio — The shape of the frame rather than its size, which decides whether output can be used without reframing and is visible only in written dimensions.
Pixel count — Width times height per frame, which drives the activation memory and explains why a figure stated at one size cannot be read at another.
Context frame — A supplied image that conditions an image-to-video run, which normally has to match the output resolution exactly, so the size binds the input too.
Latent space — The compressed representation a model actually generates in, which is why frame dimensions do not map linearly onto memory or compute.
Upscaling — Enlarging generated frames afterwards, which is a separate stage outside the release and therefore outside the resolution column.

5Terms behind the length column

Frame count — How many frames a run produces, which becomes a duration only when a frame rate is published beside it, and several releases here publish neither.
Frame rate — Frames per second of playback, which turns a frame count into a duration and decides whether output needs conversion before editing.
Clip length — How many seconds one generation produces, which decides whether a scene can be made in one call or has to be assembled from several.
Frame interpolation — Generating intermediate frames afterwards to raise the playback rate, which is a separate model and therefore outside the length column.
Temporal drift — The gradual loss of consistency across a long generation, which is why a stated frame ceiling is not the same as a usable clip length.
Sequence packing — Processing a long frame sequence in sections so the whole clip never sits in memory at once, which is how a large model can state a small floor.

Where a term is only meaningful against a real release, the note names the column and sends the reader to the statements rather than inventing an example.

Also in the register: Models, Hosted, Statements, Contrasts, Fields, Questions, Learn, Data. What a release has to publish is set out on the reading page.