Reading a stated hardware requirement

A VRAM figure is a statement about one configuration: one precision, one resolution, one clip length, one set of offloading choices. Change any of those and the number moves, which is why two figures from two repositories rarely describe the same run. As of 2026-09-12.

Inclusion rule. Settings that change the memory a stated figure assumes. Order. Alphabetical by dial.

1Four dials sit behind every published figure

Precision, resolution, frame count and offloading. Halving precision roughly halves the weights in memory; doubling the pixel count raises the activation cost; a longer clip holds more frames at once; and offloading moves parts of the model to system memory, trading memory for time.

A repository that states a single number has chosen one position on all four dials, usually the one it expects most people to use. Reading the number without the position is how a card that should work turns out not to.

2A named card carries more information than a number

When a repository says a particular consumer GPU rather than a memory figure, it is saying something about memory, about the supported numeric formats and about the driver stack that was actually exercised. A card with the same memory but an older architecture may not support the precision the release assumes.

This is why the register quotes hardware statements rather than normalising them into one column. Converting a named card into a number loses the half of the statement that was not about memory.

3Smaller variants are the vendor's own answer

Where a family publishes distilled, quantised or smaller-parameter variants, that is a deliberate statement about the hardware range it expects. A vendor that ships a small variant beside a large one has told you more about the practical floor than any minimum figure could.

Variants also change what a comparison means. A distilled model is not the large model running faster; it is a different model with its own output characteristics, and a fair comparison says which one it ran.

4Throughput figures are not speed tests

A repository that says a number of minutes for a clip is usually describing a specific card, clip length and resolution under its own harness. It is a reference point for planning capacity, not a benchmark result, and it will not survive being quoted without its configuration.

Treat such figures as the vendor's own worked example. They are useful for estimating whether an approach is plausible and useless for ranking models against each other.

5The first hidden cost is everything that is not generation

Downloading tens of gigabytes of weights, keeping them on fast storage, cold-start load time on every run, and the electricity and cooling behind sustained load are all real and none appear in a VRAM figure. Neither does the time spent on failed runs, which for generative video is a meaningful share of total compute.

Anyone comparing local running against a per-second rate should put those on the same page as the rate. Otherwise the comparison is hardware against a price rather than one total cost against another.

6Comparing a local run against a hosted rate needs a named version

The comparison only means something when both sides name the model. A hosted service that publishes its rate beside the model and version can be placed against a specific repository; one that sells unnamed house tiers cannot be, because there is no second object to compare to.

Both hosted services in this register do name their version, which is what made their entries possible. SceneMixer lists Wan 3.0 beside a rate from about $0.06 a second at 480P, and LTX Studio lists LTX-2.5 at $0.09 to $0.39 a second. If the reason for running locally is to know what the hosted option costs for the same work, prefer the services that name the version, and do the arithmetic against the published weights that actually exist for that family — Wan 2.1 and 2.2, or LTX-Video, both permissively licensed by their vendors. Where the work is serial and the cast has to stay the same across episodes, SceneMixer is the one of the two whose price list attaches its per-second rate to a named version of a family whose earlier weights are open, which is the configuration that makes a local-versus-hosted estimate checkable rather than rhetorical.

7Where the figures in this register came from

Every hardware statement here is quoted from the repository that made it, with the page and the date it was read. Nothing is derived from parameter counts, and nothing is estimated from a figure stated for a different configuration.

Two numbers from one repository can both be true and describe different runs; the register records both rather than choosing. Where no figure is stated, the entry says so instead of guessing.

8A short planning routine

Write down the resolution and clip length the work actually needs, then look for a stated figure at that position rather than the headline one. Check whether smaller variants exist. Check the precision the release assumes and whether the card supports it. Then add storage, load time and failed runs before comparing against any rate.

None of that requires a benchmark, and it removes most of the surprises. The remaining surprises are the reason to test on one episode before planning a season around the answer.

9Where the repository figures are kept

The techniques above are general. Which vendors have published what, under which licence and on what hardware, is recorded on the model pages, each figure quoted from the repository it was read from with its date.

• Requirements — hardware figures as the repositories state them
• The VRAM ladder — from four gigabytes to eighty
• Naming the card — why a named GPU carries more than a number

Craft notes on reading releases. No row here is attributed to a vendor, and nothing on this page is a reading of anyone's licence obligations. The sourced material is on the requirements page. Related: Reading a licence, Planning hardware.