Nothing in this lesson spends money. Listing providers, listing models, quoting a shot and reading your workspace are all free. So there is never a reason to guess between two candidates, quote both on the same shot and read the answer.
The quote is also the compatibility contract
No degradation is ever applied that was not named in a quote.
A quote returns a price, and it returns what that model will change about your shot. The second half is the one worth reading, and it comes in two kinds that are two different decisions.
| Kind | Meaning | What to do |
|---|---|---|
| Degrades | Still your shot, differently: five references become three, five seconds become six | Decide whether that is acceptable, then accept it explicitly |
| Blocks | The reason the shot exists is gone | Pick another model. There is no override and there will not be one |
A talking head whose creator reference would be dropped is the wrong model, not a shot somebody consents to.
Quote the shot you actually mean
A quote prices a hypothetical shot rather than reading one off the page, so it can only report incompatibilities with what you describe. Leave out the references and the one incompatibility that costs a real take cannot be reported: a creator plate handed to a model that takes no images, dropped silently. You get a clean quote for a render that will not carry the face you asked for.
That is the difference between a quote and a useful quote. Pass the references, the continuity setting, and, where a clip drives the shot, its length and frame rate, because some rows bill in frames: the same clip costs twice as much at 60fps as at 30, and without the rate they quote unknown rather than guessing.
A marketplace’s capability is a union
The single most expensive misreading in this area. Gateways run many labs’ models off one stored key, so when one reports what it supports, it reports the union across the whole shelf: the longest duration anything reaches, every resolution tier anything publishes, the highest reference count on the list.
Nothing you can actually run does all of it. A marketplace’s headline capability is a promise no single model keeps. Ask per model, and quote with the model named, that narrows to the row’s own limits before it prices anything.
One lab is not one model, either. A single vendor key can run two generations of the same family that differ by a factor in price and by a spread of seconds and resolution tiers in what they accept.
Direct or gateway: three real consequences
None of them is “marketplaces are dearer”.
- The same model sits on several shelves at different prices, and the ordering is not consistent. One gateway carries a video model at roughly three times the vendor’s direct rate, and carries another at about half what the vendor charges directly. Both are true at once, which is why comparing by reputation does not work and comparing by quote does.
- An uncatalogued model still submits, and cannot be quoted. Paste a slug we have no row for and a gateway will send it, but the estimate comes back unknown rather than invented, because a wrong price corrupts the dry run, the kept-versus-discarded meter and your spend cap all at once.
- Whether the bill is ever measured is a property of the provider. Some report what a job actually cost and your spend is reconciled against it. One large gateway returns no per-request cost at all, so every row on it is billed on our estimate forever.
The axes that actually decide
A model changes a shot it cannot make as asked. These are the changes, and each of them is named in a quote before you spend.
| Axis | What a mismatch does to your shot |
|---|---|
| References | A model that takes none drops them all. The face is not in the render, and it succeeds anyway. Over the cap, the first few are sent and the rest stay behind |
| Duration | A fixed set snaps to the nearest and is billed at the nearest. One provider accepts 4, 6 or 8 seconds and refuses 5, quoting a range that includes it |
| Resolution | A tier the model does not offer is quietly downgraded, and billed at the tier that ran |
| Aspect | An unsupported shape becomes the nearest of the same orientation. One provider has no aspect parameter at all |
| Audio | Some models always score and cannot be silenced, the track arrives and is billed whether the shot asked for one or not. Others make none, so a written line goes unspoken |
| First frame | A chained shot cannot render on a model that will not start from another shot’s last frame |
Seeds, and why variants fail on a working key
Variants are the same shot rendered several times, differing only by seed. One provider refuses a seed outright, which fails every variant batch on it. Another drops it silently, so the takes differ but none can be reproduced.
If you want a reproducible re-roll, that alone decides the provider, and it is worth checking before queueing four jobs rather than after.
Choosing, by what the job is
| The job | The axis that decides |
|---|---|
| A talking head, or any shot carrying a face | Whether it takes reference images at all |
| A shot that continues the one before it | Last-frame chaining |
| A product across several angles | The reference cap on that model, not on the marketplace |
| Dialogue or sound inside the clip | Whether it makes audio, and whether it can be turned off |
| A strict 9:16 or 4:5 delivery | The aspect list, several providers refuse 4:3, 3:4 and 4:5 outright |
| “The cheapest one” | Nothing, until you quote it |
A model that fixes a take is a different shelf from one that makes one. Do not assume the provider a lot renders on can also edit it.
Render one shot before you render the lot
The expensive mistake in this product is not choosing a bad model. It is applying a bad model to six shots at once.
One take on the candidate answers what no capability list can, which is whether it looks right. Then queue the rest.