Paper summary: Kling-Omni Technical Report

Problem

Video generation, video editing and “intelligent” reasoning-based tasks are usually handled by separate models or pipelines.

Method

Kling-Omni takes an end-to-end view. It accepts text instructions, reference images and video context, processes them into a unified multimodal representation, and generates or edits video. A comprehensive data system supports this, along with efficient large-scale pre-training and inference optimisations.

Key results

From the abstract: the paper reports exceptional capabilities in in-context generation, reasoning-based editing and multimodal instruction following, and positions the model as a step toward “multimodal world simulators”.

Limitations

There’s no dedicated limitations section in the extracted text. The report is from the product team and the abstract gives no specific benchmark numbers.