Kuaishou Kling Omni Will Be Released This Week: Achieve Director-Level Precise Control and Generate 2-Minute Long Videos with Native Audio

Kuaishou Technology's Kling AI has officially launched "Kling Omni Launch Week" this week, marking the global release of its latest multimodal video foundation model **Kling O1 (short for Omni One)**. This innovative product aims to build an "all-around creative universe" by seamlessly integrating video generation and editing through a unified multimodal vision-language (MVL) framework.

The core of Kling Omni lies in providing "director-level" control capabilities. It supports multi-input references such as text, images, and videos, allowing users to make precise adjustments to video content, such as "keeping the main character's appearance, changing the golden hour lighting, and removing background vehicles."

Compared to previous products, Kling O1 has made significant improvements in the following areas:

Duration and Coherence: It can generate continuous videos up to 2 minutes long and enhances motion stability and style coherence.
Native Audio Synchronization: It integrates native audio synchronization technology, ensuring frame-level accurate matching between sound and visuals.

The Kling series of models has already served 45 million users and 20,000 enterprises worldwide. This release has been regarded by industry experts as a revolutionary leap in the AI video field, signaling that AI video creation will shift from "random generation" to a "programmed directing" mode. This will greatly assist the efficient creation of the film, advertising, and gaming industries and further promote the deep penetration of generative AI into professional applications.

Beijing Programmers Stay Up All Night to Debug: Apple's Paper Exposed with 30% Benchmark Data Errors, ICLR Submission Undergoes Urgent Correction

Apple's vision reasoning paper submitted to ICLR 2025 claimed to surpass GPT-5, but was exposed by researchers who replicated the study, revealing serious issues: the official code lacked an image input module, and after fixing it, the accuracy dropped sharply;抽查 found that 30% of the annotated data had errors. The author team hastily closed issue reports on GitHub before finally admitting flaws in the data generation process. This incident exposed flaws in the paper review mechanism and raised concerns in the academic community about the reproducibility of AI research. (140 characters)

97% of people cannot distinguish AI music, the truth is not as bad as imagined

A study by Deezer and Ipsos showed that 97% of participants could not tell the difference between AI-created and human-made music. However, the testing method raises questions: participants were considered unable to distinguish if they got just one wrong. Although the results have attracted attention, they may overestimate the persuasiveness of AI music.

Using Poetry to Crack AI Security Defenses, Research Reveals Potential Vulnerabilities

A study in Italy found that the unpredictability of poetry may become a security vulnerability in large language models. The research team tested 25 AI models with 20 Chinese and English poems containing malicious instructions, and the results showed that 62% of the models failed to identify the hidden instructions, generating harmful content. This indicates that current AI security protection has vulnerabilities and needs to enhance content recognition capabilities.

Kuaishou Kling Omni Will Be Released This Week: Achieve Director-Level Precise Control and Generate 2-Minute Long Videos with Native Audio

Related Recommendations

Beijing Launches 'Artificial Intelligence Industry White Paper (2025)' with Expected Core Output Exceeding 450 Billion Yuan

Beijing Programmers Stay Up All Night to Debug: Apple's Paper Exposed with 30% Benchmark Data Errors, ICLR Submission Undergoes Urgent Correction

97% of people cannot distinguish AI music, the truth is not as bad as imagined

Pre-training Stuck: SemiAnalysis Leaks that OpenAI Has Not Successfully Run a New Frontier Model for Two and a Half Years

Using Poetry to Crack AI Security Defenses, Research Reveals Potential Vulnerabilities