Kling Avatar v2

Kuaishou's talking-avatar model: hand it one portrait and one audio track and it animates the face to speak that audio, with mouth shapes, expression and head motion following the recording. It is the middle rung of the three models that do this job. P-Video Avatar is a fifth of the price and reads as a photo that talks; OmniHuman 1.5 costs three times as much and moves the whole body; this one buys noticeably better facial detail and lip timing than the cheap option without the film-grade price.

The image and the audio are both required and the prompt is not, because the two inputs already carry the specification. What the prompt is for is direction rather than script — an action, an emotion, a camera move. The words are the audio's, always, and typing them again changes nothing.

It is the only model in this group that is not restricted to photographs of people: animals, cartoons and stylised characters all animate, which makes it the one to reach for when the presenter is an illustration or a mascot rather than a face. Duration follows the audio up to fifteen seconds, six aspect ratios, and one quality: 720p, at twenty-two credits per five seconds. The provider sells a second tier at twice the rate and this catalog does not offer it — the difference buys facial detail, and the model already wins that comparison against the cheap option at the price it charges here. Front-facing portraits with the face unobstructed and clean, evenly-levelled audio are what it rewards; an extreme angle or a noisy recording is where it stops holding.

Kling Avatar v2 takes the same portrait-plus-voice pair as OmniHuman 1.5, with an optional prompt that is often best left empty. It is the pick when the presenter is an illustration, a mascot or an animal. It renders at 720p and wants a portrait at least 300 pixels across.

Read the full OmniHuman 1.5 prompting guide