SkyReels-V4 Tops Global Text-to-Video Ranking

🔑 Enhanced Key Takeaways

• •SkyReels-V4 achieves an Elo score of 1131 on the Artificial Analysis Text-to-Video with Audio leaderboard, leading ahead of Kling 3.0 1080p (Pro) at 1097.[3]
• •The model supports 1080p resolution at 32 FPS for sequences up to 15 seconds, with native audio-visual co-generation including frame-perfect lip-syncing and SFX alignment.[2]
• •Kunlun Tech (also known as Skywork AI) employs a dual-stream Multimodal Diffusion Transformer architecture, starting from low-resolution text-to-image training on 3 billion images and scaling to 1080p with multimodal inputs.[4]

🛠️ Technical Deep Dive

• •Utilizes a dual-stream Multimodal Diffusion Transformer (MM-DiT) that unifies video and audio synthesis, inpainting, and editing in a single framework.[4]
• •Training pipeline: Stage 1 low-resolution text-to-image on 3 billion images, progressing to stage 6 with 1080p multimodal inputs; integrates sound via upsampling and fusion with video sparse attention.[4]
• •Supports multimodal inputs including text, image, and mask references for pixel-level control; generates 1080p at 32 FPS up to 15 seconds with microsecond-level audio-video synchronization.[2]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline