InfiniteTalk's 40 s single clip vs a 60 s Sume avatar video

InfiniteTalk is an Apache 2.0 audio-driven video model that talks about "unlimited-length" generation. Its README lists 480P and 720P output, and says single-clip mode tops out near 40 seconds (about 1000 frames), while streaming mode has no fixed cap. It is built on Wan2.1-I2V-14B and supports more than one person in a shot.

If your clip is a single presenter and under a minute, you may not need that headroom. I read the InfiniteTalk README on 2026-10-04 and compare it with the Sume avatar video limits in the docs.

What the two systems limit

When 60 seconds is enough

Most short-form placements sit inside 60 seconds, which is the ceiling for POST /v1/avatar-1.0/talking-video. You send a script (or a video_inputs scene list), and Sume estimates the spoken duration from the text. If the estimate falls outside 4 to 60 seconds the request is rejected, so you learn the problem before spending. Quality is standard, plus (the default) or max, and aspect ratio is 1:1, 3:4, 9:16 (the default), 4:3 or 16:9.

What it costs per second

Rates with no product image are $0.184 per second at standard, $0.245 at plus and $0.55 at max. A full 60 second clip at plus is therefore 60 x $0.245 = $14.70. If the script is long, split it into two requests and keep the same avatar handle, so the face stays identical between parts.

• Standard: $0.184 per second, 60 s = $11.04.
• Plus: $0.245 per second, 60 s = $14.70.
• Max: $0.55 per second, 60 s = $33.00.

When InfiniteTalk is the better fit

Choose a self-run model when you need several people talking in one frame, when a single take must run past a minute, or when you already hold the GPU capacity. Choose Sume when the shot is one presenter, the script is under 60 seconds, and you would rather approve a still preview first: previews are independent of quality tier, so you can approve once and then pick the tier at generation time.