Skip to content
Help Build a Better ComfyUI Knowledge Base Become a Patron
NewsKunlun Wanwei Open-Sources SkyReels-A2: Commercial-Grade Video Generation Framework
NEW|

Hi3DGen: A New Framework for High-Fidelity 3D Geometry Generation through Normal Bridging

04/07/2025

Kunlun Wanwei Open-Sources SkyReels-A2: Commercial-Grade Video Generation Framework

On April 6, 2025, Kunlun Wanwei officially open-sourced its latest developed SkyReels-A2 model, the world’s first “Elements-to-Video” (E2V) generation framework aimed at commercial scenarios. Through an innovative dual-branch architecture, this framework can transform multiple reference images into coherent and fluid video content, marking the transition of AI video generation technology from experimental stage to practical application.

SkyReels-A2 Demo

Technical Highlights: Dual-Branch Architecture Breaks Through Video Generation Bottlenecks

The core innovation of SkyReels-A2 lies in its unique dual-branch feature encoding system:

  • Spatial Feature Branch: Uses a refined VAE encoder to process images, extracting texture and detail information of characters, objects, backgrounds and other elements, ensuring high consistency between each element in the generated video and the reference images.

  • Semantic Feature Branch: Utilizes CLIP visual encoder and MLP projection layers to capture high-level semantic associations between elements, integrating them into the diffusion model through cross-attention mechanisms to ensure logical coherence and dynamic continuity of scenes.

This design successfully solves technical challenges in multi-element consistency control and complex scene semantic coordination that traditional video generation models face, making the generated videos more fluid and realistic than many closed-source commercial models.

Broad Applications: Comprehensive Empowerment from E-commerce to Film Production

SkyReels-A2 demonstrates powerful application potential in multiple domains:

  • Virtual E-commerce: Simply input a host image and product pictures to generate dynamic recommendation videos, solving the problems of high cost and long production cycles of traditional advertising.

  • Film Production: Supports combinations of multiple characters and backgrounds, capable of generating movie-level interactive scenes such as group escapes in disaster films or character interactions in dramas, with composition and lighting effects reaching professional standards.

  • Music Multimedia: Can combine background elements and rhythms to generate music video segments, providing independent musicians with low-cost creative tools.

Open Source Ecosystem: Promoting Industry Technology Accessibility

This open-source release is an important step in Kunlun Wanwei’s AI video sector strategy. Previously released SkyReels-V1 (short drama generation model) and SkyReels-A1 (expression and action control algorithm) have already accumulated a large developer ecosystem. SkyReels-A2 further provides:

  • Efficient Inference Framework: A single RTX 4090 GPU can generate 544p video in 80 seconds, with support for multi-card parallel processing and low VRAM optimization.

  • Structured Data Processing Pipeline: The entire workflow from video annotation, element segmentation to triplet matching is open-sourced, significantly lowering the application threshold for enterprises.

Model Specifications and Technical Parameters

SkyReels-A2 offers multiple model versions to meet the needs of different application scenarios:

  • A2-Wan2.1-14B-Preview (Released): Supports generation of approximately 81 frames at 480×832 resolution
  • A2-Wan2.1-14B (Coming Soon): Base version with video parameters same as Preview version
  • A2-Wan2.1-14B-Infinity (Coming Soon): Supports generation of unlimited length videos with increased resolution of 720×1080

The model is based on a video diffusion transformer architecture, using an innovative dual-branch encoding system to achieve precise control over reference images, ensuring high consistency of objects, characters, and background elements in generated videos.

Recent Development Plans

The Kunlun Wanwei team has announced recent development plans for SkyReels-A2:

  • Launching the A2-Bench evaluation system and leaderboard
  • Releasing the complete model sequence, including versions supporting unlimited length video generation
  • Optimizing inference performance for RTX 4090 GPUs
  • Integrating ComfyUI support, making it easier for users to use the model through a graphical interface

Industry Impact and Future Outlook

The release of SkyReels-A2 fills the gap in commercial-grade control capabilities of open-source video generation models and may change traditional video production processes. Industry experts believe this technology will accelerate the popularization of personalized content production and real-time interactive media, such as generating live-streaming e-commerce videos by combining real-time motion capture, or dynamically building virtual environments for metaverse scenarios.

The Kunlun Wanwei team states they will continue to optimize the model’s capabilities in long video temporal consistency and physical engine interaction, and explore deep integration with 3D modeling tools.

Related Links