Kling (1.0)

What it is

Kling (可灵大模型 1.0) is the first version of the video generation model from Kuaishou Technology (快手科技). It generates video from text descriptions, and later versions added image-to-video and video extension.

Origins

Kuaishou launched Kling's text-to-video generation on 6 June 2024 and opened applications for testing on the same day.^[1][2] The first testing was offered inside KuaiYing (快影), a video-editing app that Kuaishou operates, and was available to users in China.^[3][4] Within a month of launch, more than 500,000 users had applied for the beta and about 7 million videos had been generated.^[1]

Architecture

Kling uses a diffusion transformer (DiT) architecture. A diffusion model learns to produce data by starting from random noise and removing it step by step. A transformer is a neural network design that models relationships across a sequence of tokens. Kuaishou combined the two into a diffusion transformer.^[3]

Kuaishou also developed its own 3D VAE (three-dimensional variational autoencoder). A VAE compresses data into a smaller representation and then reconstructs it. Kling's 3D VAE compresses space and time together, so a video is handled as a block of frames rather than as separate images. A full-attention spatiotemporal modelling module lets the model consider every part of a video across both space and time.^[3]

Main features

Kuaishou said Kling could generate videos of up to two minutes at 30 frames per second and up to 1080p resolution, in a range of aspect ratios.^[3] The company designed the model to simulate behaviour in the physical world, including fast-moving objects, sudden changes in a scene, and complex human movement.^[3]

The model's feature set grew after the first release. Image-to-video and video extension came next. At the World Artificial Intelligence Conference (WAIC) on 6 July 2024, a third upgrade added a web version, HD output, start-frame and end-frame control, camera control, and text-to-video clips of 10 seconds.^[1]

Reception

TechCrunch reported in July 2024 that Kling appeared to censor politically sensitive topics.^[5] The report is one of the early public observations about how a Chinese video generation model handles sensitive content.

Successor

Kling was succeeded by Kling 1.5 in September 2024.^[6]

Its place in the field

Kling 1.0 is the starting point of the Kling line. Its launch made a video generation model available to users through a China-only beta and then a global beta. The global beta opened on 25 July 2024 at klingai.com, after more than 1 million applications.^[2]

For a student of AI, Kling is a useful example of how a video generation model is built and released in stages: first a limited beta, then new features, then a global release, with each step adding capability and user reach.^[1][2]