Which engine for which job

Switching model costs nothing

A graphic costs one credit whichever engine renders it, and the model is a single click before you render. So there is no reason to choose once and for all, or to stay with one because "it worked last time". Video is different: there the per-second rate varies several times over, and the choice has a real effect on how many clips you get.

Four questions that narrow the list

• Does the frame have to show your exact item? If so, you need an engine that takes reference photographs — nine of the sixteen image engines do.
• Is there text on the graphic? If so, the list shortens to the few models that set letters correctly.
• What is the format? Only some engines manage very wide or very tall frames; several take ratios from a closed list rather than an exact pixel size.
• Is this the version you will publish, or an attempt to test an idea? That decides whether you pay in time or in quality.

The answers to the first three are in the spec table on each model page — the number of reference photographs, the ratios, the sizes, the cost. That table is filled from the same registry the app itself uses, so it cannot drift away from what the engine will actually do.

Search cheaply, finish expensively

This is the most profitable habit in the whole app. Engines range from six seconds to forty for an image, and from one minute to four for video — and at the stage where you are testing whether a composition works, quality is beside the point, because you will reject three versions out of four anyway.

1. 1Settle the composition in the fastest engine: FLUX.1 Schnell or Z-Image Turbo for graphics, PixVerse V6 for video.
2. 2Run exactly the same brief through the model you actually want. Do not rewrite it — the engine should be the variable, not new wording.
3. 3Only then order the versions and formats you will publish.

With video the difference is countable. An eight-second clip from Veo 3.1 is nine credits, and four seconds from PixVerse V6 is one. Every rejected attempt put through the cheaper engine rather than the target one is eight credits you keep, and rejected attempts are the rule rather than the exception.

When the problem is the model, not the brief

A third failed attempt with the same brief almost always means the engine simply does not do that. Rewriting the brief a fourth time is a waste of credits at that point — the model has to change, not the sentence.

• You get a different frame shape than you ordered: the model takes ratios from a list and yours is not on it. Move to one that takes a size in pixels.
• The lettering comes back with a typo: that is the engine, not the brief. Move to a typography model.
• Invented text appears on the label although the brief says to leave it blank: a known habit of several models, and not something a brief can switch off.
• The clip has no sound, or the product is not yours: some video engines generate no audio track and do not start from a photograph. That is in the table, not in the brief.
• A thin loop — a mug handle, a bag strap — did not survive into the 3D model: two of the five engines cannot do it, and higher detail does not fix it.

Every model page carries a "When to pick something else" section that says plainly where that engine is weak and names a specific alternative. That is usually the shortest route to a result.