Kunlun TianGong SkyReels-V4 Ranks Second Globally in AI Video Generation

Kunlun TianGong's SkyReels-V4 has secured the second position among active models on Artificial Analysis's global text-to-video leaderboard, which includes audio generation. The Chinese-developed model, which also ranks fourth among all historical models, entered the top tier alongside systems such as Veo 3.1, Sora 2, Kling 3.0, and Wan 2.6.

Digital leaderboard interface showing SkyReels-V4 ranked second among AI video generation models.

Artificial Analysis's leaderboard is recognized by leading AI companies including OpenAI, Google, and Anthropic.

Advanced Video Generation Capabilities

SkyReels-V4 is designed to function as a video foundation model that supports multimodal input, joint audio-visual generation, and unified generation, repair, and editing tasks. The system can produce 1080p, 32 frames per second (FPS), 15-second cinematic videos from minimal inputs, such as two character images and dialogue.

Video editing timeline on a screen, showing synchronized audio and video tracks for two characters.

The model demonstrates precise audio-visual synchronization, including lip-sync, even when processing multiple reference images and audio inputs for diverse characters. Its multimodal referencing capability allows users to combine inputs like character designs and dance reference videos to generate new content, maintaining character consistency and motion quality.

SkyReels-V4 also features a first-frame reference function for flexible multimodal control, enabling the creation of varied scenes from initial images. A motion reference function allows for consistent movement transfer between different character styles.

Integrated Repair and Editing Functions

The platform offers comprehensive repair and editing functionalities. It can remove watermarks, delete specific characters from scenes using prompts, and alter character clothing with a single click. Additionally, SkyReels-V4 can transform an entire video's style, for instance, from realistic to a Lego-brick aesthetic, while preserving movements, camera work, and plot rhythm. This capability allows for rapid visual language switching to meet commercial demands and supports IP secondary creation and brand collaborations.

Visual metaphor of a video transforming from a realistic style to a Lego-brick aesthetic.

The system can also convert static shots into dynamic cinematic movements such as push-ins, pull-outs, pans, and tilts, enhancing narrative tension. Tasks that typically require multiple professional software tools can be managed within SkyReels-V4.

Core Technological Innovations

SkyReels-V4's performance is underpinned by three key technological advancements, detailed in a technical report.

The first is a dual-stream Multi-Modal Diffusion Transformer (MMDiT) architecture, which integrates audio and video processing. Unlike models that generate video and then attach audio, SkyReels-V4 uses a symmetrical dual-stream design where video and audio branches interact deeply through a bidirectional cross-attention mechanism. This ensures that video and audio generation are co-evolving. The system also employs RoPE frequency scaling to align audio positional encoding frequency with video frame rates, achieving millisecond-level temporal alignment.

The second innovation is a unified splicing framework that consolidates various video operations into a single model. This framework uses a "channel splicing + temporal splicing" paradigm, treating tasks like text-to-video, image-to-video, video extension, regional repair, and style transfer as repair problems. By adjusting mask configurations, the model can seamlessly switch between tasks, allowing users to express their creative intent directly.

Third, an efficient generation strategy balances quality and speed. To handle 1080p resolution, 32 FPS, and 15-second durations, the engineering team developed a "low-resolution full sequence + high-resolution keyframe" joint generation strategy. This involves generating a low-resolution video and high-resolution keyframes, then reconstructing the final video using super-resolution and frame interpolation modules. The self-developed Video Sparse Attention (VSA) mechanism reduces attention computation costs by approximately three times. The model also uses a multi-stage progressive training strategy, starting with 256px text-to-image pre-training and gradually expanding to higher resolutions.

Abstract visualization of dual-stream data flow, representing integrated audio and video processing.

Synergy with AI Music and Future Directions

SkyReels-V4 is part of Kunlun TianGong's broader "full-modal AI creation ecosystem," which includes Mureka, an AI music creation platform. Mureka O1 is noted as the first large music reasoning model to incorporate Chain of Thought (CoT) technology, with Mureka V8 further improving timbre, playing techniques, and emotional expression.

The combined capabilities of SkyReels-V4 and Mureka allow for integrated visual and audio content creation. This enables the generation of complete video advertisements from a single text prompt, high-quality music videos from songs, and educational videos with explanatory audio and dynamic visuals.

While other companies like Google, Adobe, and ElevenLabs are also exploring integrated video and music creation, Kunlun TianGong has demonstrated this synergy through products like an AI musician MV, where Mureka generated the music and SkyReels provided the video.

Artistic representation of video and music elements merging, symbolizing AI creative synergy.

The emergence of SkyReels-V4 coincides with a dynamic period in the AI video generation market. While Seedance 2.0 has gained attention, SkyReels-V4 offers a robust alternative, particularly for users requiring stable, professional, and comprehensive video creation capabilities.

Continuous Iteration and Ecosystem Development

Kunlun TianGong's development trajectory for large video models includes several iterations:

• February 2025: SkyReels-V1 open-sourced, focusing on AI short drama creation with film-grade texture.

• April 2025: SkyReels-V2 released, a model for infinite-duration film generation.

• January 2026: SkyReels-V3 open-sourced, introducing multimodal video generation with support for multiple reference images.

• February 2026: SkyReels-V4 released, achieving its current global ranking with multimodal input and unified task framework.

This iterative approach, combined with Mureka's advancements in AI music and the Skywork series in large language models, contributes to Kunlun TianGong's AI ecosystem, which spans computing power, models, and applications.

Kunlun TianGong's technical report outlines future directions, including extending video generation to longer durations (30 seconds+), enhancing real-time interactive editing, and integrating model APIs with creative tool ecosystems.