What is SkyReels V4
SkyReels V4 is a video foundation model launched by Kunlun Wanwei, and the world's first AI video model supporting multimodal input, joint audio-video generation, and unified generation/restoration/editingAI video model. The model uses a dual-stream MMDiT architecture and can generate 1080p/32FPS/15-second cinema-grade synchronized audio-video. It ranks first on the Artificial Analysis global leaderboard in the Text to Video (With Audio) track, surpassingGoogle Veo 3.1、OpenAI Sora 2and other mainstream models, and supports multimodal control via text, images, video, and audio as well as professional-grade video restoration and editing.SkyReels V4 has officially launched its API entrance, opening API capabilities forall scenarios.
Main features of SkyReels V4
- Precise multimodal control: Supports various input combinations such as text, images, video clips, masks, and audio references to achieve subject appearance preservation, timbre transfer, and motion replacement.
- Professional-grade video restoration: Through region-based intelligent restoration and reference-guided restoration, it precisely replaces video subjects, modifies attributes, or changes backgrounds, ensuring visual consistency before and after editing.
- Full-dimensional video editing: Supports local editing (adding or removing objects, modifying textures), intelligent element removal (watermarks/subtitles/logos), as well as global style transfer and scene attribute adjustment.
- High-quality audio generation: The model has built-in multilingual speech synthesis, sound effect generation, and background music adaptation, supports emotional speech and synchronized singing with lyrics, and performs outstandingly in Chinese speech.
Technical principles of SkyReels V4
- Dual-stream MMDiT architecture: It adopts a symmetric dual-stream design, with the video and audio branches sharing the MLLM text encoder, achieving deep audio-visual synchronization across the entire network through a bidirectional cross-attention mechanism; it uses RoPE frequency scaling technology to solve the mismatch in audio-video time scales, and combined with a joint flow matching loss function, fundamentally solves the challenges of lip synchronization and sound effect alignment.
- Unified concatenation framework: It innovatively introduces a two-dimensional paradigm combining channel concatenation and temporal concatenation, unifying diverse tasks such as generation, restoration, and editing into restoration problems under specific mask configurations, achieving one-stop coverage of all-scenario video operations and enabling end-to-end creation without switching tools.
- Efficient generation strategy: The model adopts a joint generation strategy of "low-resolution full sequence + high-resolution keyframes", and with a video sparse attention mechanism, reduces attention computation cost by about 3 times, making 1080p high-resolution long-duration video generation practically valuable.
Project address of SkyReels V4
- Project official website：SkyReels
- arXiv technical paper：https://arxiv.org/pdf/2602.21818
Application scenarios of SkyReels V4
- Advertising and marketing: The model can quickly generate product promotional videos, supports multi-style switching and batch editing, and improves advertising production efficiency.
- Content creation: The model supports visualization of short video scripts, intelligent editing and restoration of Vlogs, and synchronized multilingual dubbing, lowering the creation threshold.
- Film and television production: Used for early-stage concept visualization, shot extension, post-production restoration, and local editing, accelerating the industrialization process of film and television.
- Education and training: The model supports teaching video generation, courseware visualization, and automatic synchronization of multilingual subtitles, facilitating online education content production.