Building a lot by hand is six steps: resolve the inputs, plan the beats, create the shots, attach the references, set the look, and price it. A recipe is those six in one call, ending with a shot list and a number in the same answer.
It still spends nothing. That is worth saying because a call that does six things sounds like it must, but planning and creating are free, and the number a recipe hands back is a quote rather than a charge.
What you get back
| In the answer | Read it for |
|---|---|
| The shot list | Whether the beats are the ad you described |
| The estimate | What one take of every shot would cost. A second take of a shot costs again |
| Any refusals | What the vertical turned down, and why |
| The overrides | Which axes of your look the capture style took. An override you discover in a render is one nobody told you about |
A dry run does all of the above and creates nothing at all, so you can see the shot list and the price for an idea you are not sure about, and then not build it.
When to use the long way instead
The plan-then-create pair exists for exactly one situation: you want to edit the plan between planning and creating. If you are going to take what it gives you and then fix a shot, use the recipe and fix the shot, changing one shot afterwards is cheap and nothing about a recipe makes it harder.
Runway’s recipes are a different thing
Credits are debited when the call is accepted, and a cancelled task is not refunded.
Everywhere else in this product, cancelling a queued job is a way out. Here it is not, the credits are gone whether or not you keep the result. So the dry run is not a nicety, it is the only chance to change your mind.
These make a whole finished piece from named inputs. No model to pick, no prompt vocabulary to learn: you supply the pictures and the recipe does the rest. Seven of them, split by one question, does this make something from nothing, or change something you already have?
The four that make something
| Recipe | You supply | The thing to know |
|---|---|---|
| A finished product ad | Product pictures | Four seconds or more |
| A creator holding your product | Who is on camera, plus one product | Vertical only. A landscape ask is refused rather than quietly turned on its side |
| Several shots from one prompt | A prompt | The duration is exactly 5, 10 or 15, not a range, not a maximum |
| Marketing stock images | A prompt | Four pictures, each a different camera angle on the same subject. The count is not askable |
That last one is worth planning for. Nothing this product can send moves the count, the price is flat for all four, and they land as four takes of the shot to choose between. Somebody expecting one picture will think something went wrong.
The three that change something
- Replace the product inside a clip. Needs pictures of the new product and a picture of the one being replaced. Supplying only the new one is the common mistake, the recipe cannot infer what it is looking for.
- Remake a still in another language, from a list of twenty-two.
- Campaign pictures from one still, plus a prompt.
Two of these have a flat price, one figure, with no field that could make it cost anything else, which makes them unusually easy to quote exactly if you are used to per-second estimates.
The results are ordinary takes
A Runway recipe runs on a shot and lands as a take of it. So it is comparable against anything the studios rendered, pickable as the winner, and exportable in the same pack. It is not a side channel with its own gallery.
Which means everything else you know still applies: you can follow the job, mount it if it wins, label it so four stock images are tellable apart, upscale it, and ship it. And you can quote a conventional render on another provider first, which is the number a recipe is competing against, worth doing, because the credits are not coming back.
The shot’s kind has to match: the video recipes need a video shot, the image one needs a stills shot. And a recipe fills one shot at a time. It does not replace rendering a whole lot.