OmniHuman vs Kling Avatar vs HeyGen: one song, six models

OmniHuman is ByteDance’s model for turning one photo and an audio track into a video of that person talking or singing. Most of the Google first page for its name is unofficial sites built on the name. We wanted a portrait to sing a song we already had, so on 2026-11-08 we ran the same input through six models: one painted portrait of a young woman and the same 5-second chorus, silenced before the first word and after the last. Kling AI Avatar v2 Standard won. The mouth stayed shut until she started singing, the shapes matched the words, and the face stayed the face in the portrait. Kling’s Pro tier cost twice as much and looked the same to us. OmniHuman 1.5 moved the head and body the most, but the face drifted furthest from the portrait, and it was the most expensive. VEED Fabric was by far the fastest, done in 26 seconds, but the mouth barely opened. HeyGen Avatar IV, through its API, was the cheapest and second fastest: 33 seconds for the clip. Its mouth shapes were close to Kling’s, but on our first try it smiled and showed teeth even before the first word. HeyGen takes its resting expression from the photo, and ours had a half smile. With a neutral, closed-mouth version of the same portrait and a two-word motion prompt, “sad, almost still”, the smile was gone. Two things surprised us. First, the prompt matters: Kling takes a text prompt for expression, and a cheerful one made our sad character grin and even added fake captions; writing “calm, slightly sad, no smile” fixed both. Second, a video model given your song as a reference does not lip-sync it. Seedance 2.0 with reference audio sang its own garbled version of the line, so the mouth followed its singing, not ours. Mouth-only lip-sync models (sync lipsync-2 and 2-pro) repaint the lips on a video you already have; on our close-up they drifted off the words. For a still portrait that has to sing, a photo-driven avatar model did better. Through the APIs, each 5-second clip cost between $0.20 and $0.80. Each clip is generated against its own audio: we reused one chorus clip for the second chorus and it was visibly out of sync, because she phrased the line differently the second time.