Which model for which job
71 models is too many to choose from. Here is how the choice actually narrows.
2 min read · last checked
Start from what you are bringing
The first split is not quality, it is input. What you have in your hand decides which half of the catalogue can even run.
- Nothing but a sentence , 17 video models start from text alone, and every one of the 15 image models does.
- An image you want changed , that is an edit model. There are 11, and they take your picture as the source instead of a description of it.
- An image you want to move , most video models work this way, from a single start frame.
- Two images, a beginning and an end , 3 models build the clip that runs between them.
- A person or product that has to stay the same across shots , 4 models work from reference images rather than one frame.
- A clip you already have , 4 models rework existing footage.
Then let the price decide
Image models sit close together, from $0.030 per picture upward, so the cost of trying one is small. Video is the opposite: every video model charges by the second, and the spread between the cheapest and the most expensive is more than a factor of ten.
That is why the studio shows the price on the button before you press it, recalculated as you change resolution, length and sound. A ten-second clip at the top resolution on the heaviest model is not a small decision, and it should not look like one.
A model with no published price is not in the catalogue at all. Providers do list such models, and the honest thing is to leave them out rather than invent a number and hope.
Then the character of the model
Within a price band the models genuinely differ, and not on a scale from worse to better. One is strongest at readable text inside the image. One holds a face steady across a series. One has a look of its own that you either want or do not.
Each model in the studio carries what it is strong at and what it is less suited to, researched with the source noted. The assistant uses the same notes when it picks for you, and it says why it picked.
And your own style on top
A LoRA is a small file trained on your own material: a face, a product, a house style. 7 of the models take one, and it sits on top of whatever else you asked for.