Griffin leads DOVER, FID, THEval: pull frames to judge a clip

Tavus says Griffin-Lite ranks first on DOVER, FID and THEval and second on LSE-C with 7.27 (Tavus Griffin page, read 2026-10-06). You can't run those metrics on a Sume clip from Sume, because Sume does not ship them. What you can do is pull exact still frames from your rendered clip with Video frames and check the things those metrics stand for by eye: image quality, a face that stays the same, and a mouth that matches the words.

What the page lists

The page ranks Griffin-Lite first on DOVER (a video quality score), FID and THEval, and second on LSE-C, a lip-sync measure. Tavus gives 7.27 for LSE-C. It does not say here how a clip you produce would score, and none of this is a test of a clip you render with Sume.

The manual version with Sume

POST /v1/video-frames takes one Sume-hosted clip and either at[] times or an fps, and returns durable image artifacts at the source size. A submit always returns 202, so poll the job and read the artifacts when it completes.

Pull a still at the first spoken word, one mid-sentence and one on the last word. Look at the teeth and lips for blur, at the hairline for flicker between stills, and at the eyes for a consistent look. This is a spot check, not a benchmark.

• First spoken word: is the mouth open on the right shape?
• Mid-sentence: does the face stay sharp and the same person?
• Last word: does the clip end cleanly, not mid-gesture?

Cost and limits

Video frames bills by its Modal compute rather than a flat price, and reserves a ceiling at submit. Read the page before looping over many clips. Avatar Video itself renders at 720p at this time.