Runway vs Luma, judged on hit rate, not samples

A reel tells you what a model does at its best. It tells you nothing about how often.

The number that governs your month is hit rate. Usable seconds per ten generations of the shot you actually need, on your subject rather than on a demo prompt.

Budget from that. At one usable attempt in ten, a six-shot sequence costs sixty generations. Teams who plan one generation per shot run out of credits in week one.

So run the same brief through both from the same reference image. Ten times each. Count what you would use.

Do it on the subject you sell. Hit rate varies with faces, hands, liquids, fabric and machinery. A tool that impresses on landscapes can fail on the one product your campaign is built around.

Time the wait as well as the result. A model that takes four minutes an attempt costs you the afternoon, not the credits.

Multiple generation modes. Image and video conditioning. Editing and clean-up. Then the unglamorous parts that show years of use: asset handling, iteration history, an interface built for forty attempts.

That pays when consistency matters. Holding a look across a sequence. Iterating from a still you control. Keeping a project findable while other people work on it.

It suits work with a delivery date attached. The tooling around the generation is what stops a deadline slipping.

The trade is weight and cost. More surface to learn, and heavy iteration burns credits quickly. Price a real week of your own work before committing to a year of anything.

Check what happens to the material afterwards. Where generations are stored and what an export gives you decides whether last quarter's work is reusable or gone.

A short loop from idea to something on screen compounds. Most attempts fail in this medium. A tool giving you three times as many shots an hour can beat a stronger model.

That suits exploration. Mood films, concept work, social-native visuals where interesting matters more than exact.

The trade shows up when the shot is specified rather than discovered. A lighter control surface bites hardest on the shot somebody asked for by name.

Test that directly. Write down the one shot in your current project you cannot fake. The one with a named product or a specific action. Generate only that, in both tools.

Easy shots stopped separating these products a year ago.

One beautiful five-second clip is achievable in both. Character, product and lighting persisting across six shots is still heavy work. An ad is six shots.

Duration compounds it. Clips are short and extensions drift. A coherent thirty seconds gets assembled from many attempts rather than generated in one.

So the common professional use is supplementary. An impossible establishing shot. A texture. A transition. Two seconds you could not film.

Budget the rest as filming. A day with a camera produces material both tools would spend a week failing to invent.

Plan the edit that way. Generated shots for the parts you cannot shoot, real footage for the rest. The budget then survives contact with the deadline.

Read the commercial terms before the first render. They differ by tier and both companies have revised them, so check what your plan permits for paid distribution.

Most performance video is a person making a claim. A believable person making a believable claim is the hardest thing either model does.

Cutroom takes the opposite route, and it is the cheapest route to that shot by a wide margin. Film one take of up to three minutes on a phone.

It comes back as a finished 9:16 MP4 you direct on the transcript. Highlight a phrase and coverage lands over those words. Delete a line and the cut rebuilds. Change the pace and it re-cuts.

Captions ship in six packs. B-roll comes from a free library and is capped by pace at 40, 45 or 52 percent, so the pictures land where the words need showing.

There is one attempt and one price. A batch is 100 credits and an export is 20 credits per output minute, so a thirty second ad is 110 credits with no rerolls to pay for.

The facts a buyer needs: it generates nothing, three minutes is enforced at upload, and output is 9:16. The trial runs 7 days on 300 credits with no card.