Kuaishou has launched Kling 3.0, a new version of its AI video generator. It adds native 4K output, sequences of several shots from one prompt, and sound generated together with the picture. CineD's coverage puts it in direct competition with OpenAI's Sora, Runway and Google's Veo, and says early users see a clear jump in photorealism.
The big change is that image, video and audio now run through one system instead of separate tools. Kuaishou calls this a Multi-modal Visual Language framework, built on a diffusion transformer. Characters made with the companion Kling Image 3.0 carry into video without their features drifting, which has been a common weakness when work passes between models. The company says detail is generated at 2K or 4K directly, not upscaled afterwards.
Kuaishou pitches this as an "AI Director" approach. A single generation can run up to 15 seconds and contain several cuts, with characters kept in consistent positions as the camera angle changes, so the output looks more like coverage than a lone clip. Prompts can ask for moves like dolly shots or rack focus, and a physics layer handles weight and momentum.
On audio, the model makes dialogue, lip sync and ambient sound in the same pass. It can tie a specific voice to each character in a scene, and it supports English, Chinese, Japanese, Korean and Spanish. An Elements feature takes reference images or clips to fix a character's look, build, movement and voice across shots. An image-series mode aims at storyboarding, and the image model now renders legible text such as signs and screens.
CineD also compares it with rivals. It reports benchmarks that give Kling an edge over Runway in prompt adherence and realistic human movement, credits Veo 3 with better lip sync, and says narrative filmmakers tend to prefer Kling's lighting and cinematic look. Unlike Sora, it can be used today on a subscription.
CineD notes some practical limits: wide cinema ratios have to be cropped from 16:9, and longer scenes mean stitching 15-second clips together. It also flags that Kuaishou hasn't said what the model was trained on, and argues that working filmmakers need to keep up with these tools and judge for themselves which ones to use.
Sources · 1