Kuaishou launches the Kling video generation model

Background

A text-to-video model takes a written description and returns a moving clip instead of a still image or a block of text. Most systems of this kind work by diffusion: the model starts from random noise and removes it step by step, guided by the prompt, until a sequence of frames appears. Video makes the problem harder than image generation because the pictures must stay consistent from one frame to the next, and because the amount of data involved grows quickly with length and resolution.

What happened

On 6 June 2024, Kuaishou launched Kling, a text-to-video generation model, and opened applications for testing.^[1][2] Testing was first available to users in China through the company's KuaiYing app.^[3] Kuaishou later said the beta was open to global users through klingai.kuaishou.com and klingai.com.^[2]

Key details

Kuaishou said Kling could generate up to two minutes of 1080p video at 30 frames per second.^[3] That combination of length, resolution and frame rate was the capability the company put forward at launch. The model is built on a diffusion transformer together with a self-developed 3D variational autoencoder, or VAE.^[3] In rough terms, the transformer handles the sequence of frames and their relationship to the prompt, while the 3D VAE compresses video into a smaller latent space and decodes it back into pixels; doing the generation in that compressed space is a common way to keep the computation for many frames manageable.

Kuaishou described Kling as the world's first accessible, real-image-level video generation large model for ordinary users.^[2] That is the company's own description of the release rather than an independent measurement.

Take-up

The figures Kuaishou reported after launch describe how quickly people tried the model. Within a month, more than 500,000 users had applied for testing and 7 million videos had been generated.^[1] In the global beta, the company said it received over one million applications and granted early access to more than 300,000 users.^[2]

Why it matters

Kling was released with a public application process rather than only as a technical report, so ordinary users could sign up and try it.^[1][2] The gap between the number of applications and the number of users let in also shows how access was staged: demand arrived faster than the company handed out testing slots.^[1][2] The global beta extended the same application route beyond China through two dedicated sites.^[2]

On the technical side, the two-minute 1080p capability and the in-house 3D VAE show that the release covered a full pipeline rather than a single component.^[3] For people watching the field, Kling is a case of a video generation model shipped as a public, testable product with stated limits on length, resolution and frame rate, and with usage numbers published by the company that made it.^[1][3]