Kuaishou announced early access to its 3.0 series for Ultra subscribers. Video 3.0 supported clips up to 15 seconds long, and Video 3.0 Omni added a storyboard where creators could set the duration, framing, and camera movement of individual shots.
Kling AI
Also known as: Kling · 可灵 · Kling Video
Plain English
Kling is Kuaishou’s family of video generation models. A creator can describe a scene in words or supply reference images to guide how a person, object, or setting looks.
Technical Overview
Kuaishou described the original Kling model as a diffusion transformer operating on video compressed by a 3D variational autoencoder across both space and time. Later versions added controls for reference subjects and individual shots, as well as joint generation of sound and images.
Chronological Entries (4)
Kuaishou released Kling Video 2.6 with joint audio and video generation. A single generation could produce the moving image, dialogue, and environmental sounds, including background noise that matched the scene.
Kuaishou reported that multi-image reference was available worldwide in Kling 1.6. A creator could supply separate pictures of two subjects and use the prompt to describe their interaction, giving the model a visual reference for each subject’s appearance.
Kuaishou opened testing of Kling’s text-to-video feature. In the original model, a 3D autoencoder compressed frames across space and time, giving the diffusion transformer a smaller representation of the clip to generate.