Five people in the clip: Recast takes four photos, 12 s costs $4.50

Sume's H3 Max Recast takes between one and four reference photos, one for each person you want to swap in, so a clip with five people on screen cannot be fully swapped in one job. The request with five image URLs is refused by the Video Router schema before any paid submission, which means you pay nothing for the failed attempt. A 12-second source swapped with four photos at 768p costs 12 x $0.375 = $4.50.

What the limit is and where it comes from

The Video Router docs describe h3-max-recast as replacing the people in a source video with the people from 1 to 4 reference images, with one photo for each person, at 768p or 1080p, for 5 to 30 seconds. The Sume code mirrors the provider's cap of four people. There is no setting that raises it, and the Sume API test suite checks that five reference images are rejected.

Options for a five-person clip

The honest choices are limited. First, check whether all five people are on screen for the whole clip; if one person only appears in part of it, you may be able to trim that part out. Sume has a video trim route at a flat $0.02 per job. Second, run two clips, with the people split across them, by trimming the source into two ranges and giving each range its own set of up to four photos. Third, change which people you swap and leave one person as filmed.

I have not verified how the model treats a person who is in the source but has no matching photo, so test that on a short, cheap clip before you build a pipeline around it. A 5-second test at 768p is the floor and costs 5 x $0.375 = $1.875.

Cost of the two-pass route

Two passes over a 12-second source, each pass the full 12 seconds, cost 2 x $4.50 = $9.00 at 768p, plus $0.02 for each trim. Two passes over trimmed halves of 6 seconds each cost 2 x 6 x $0.375 = $4.50 plus the trims, but only work if the groups of people do not overlap in time. Do the arithmetic against your own cut list before you commit.

A decision table

Use the table to pick a route before you spend money. The costs assume a 12-second source at 768p and use the Sume figures above.

What I could not confirm

The docs say one photo for each person and a maximum of four. They do not say what the model does with a fifth person who has no photo, and they do not promise that the same person keeps the same identity across two separate passes. If your edit needs a consistent look across two jobs, use the same photo for the same person in both and compare the first frames of each result before you join them.

Cost of the test: a 5-second trial at 768p is $1.875, and two of them are $3.75. That is the price of finding out, and it is cheaper than a full second pass.