On February 18th, Kunlun Tech unveiled China’s first AI short film creation video generation model, SkyReels-V1, and China’s first SOTA-level emotion-action controllable algorithm based on a video pedestal model, SkyReels-A1.
SkyReels-V1 is an open-source video generation model dedicated to AI short film creation, capable of generating cinematic-level character micro-expressions and movements. Not only is SkyReels-V1 one of the few global open-source video generation models, but it is also among the strongest in terms of performance in open-source video generation models, supporting 33 delicate character expressions and over 400 natural action combinations, highly replicating human emotional expressions and possessing movie-level aesthetic lighting. Trained on high-quality film and television data, it supports text-based video and image-based video functions, leading in parameter scale among open-source models.
From a technical perspective, with the support of the self-developed inference optimization framework “SkyReels-Infer,” it significantly enhances inference efficiency, achieving 544p resolution with inference based on a single 4090 taking only 80 seconds, supporting distributed multi-card parallelism, Context Parallel, CFG Parallel, and VAE Parallel. Additionally, it adopts fp8 quantization and parameter-level offload to meet the operational requirements of low VRAM user-level graphics cards; supports flash attention, SageAttention, model compilation optimization, and further optimizes latency; and enhances usability based on the open-source diffuser library.
SkyReels-A1 is China’s first SOTA-level emotion-action controllable algorithm, supporting video-driven movie-level emotion capture and high-fidelity micro-expression restoration. This algorithm can generate realistic dynamic character videos based on any human body proportion, supporting side-face expression control and eyebrow-eye micro-expression generation, and enabling more extensive character action driving. Compared to similar algorithms, SkyReels-A1 excels in the complexity and naturalness of emotion-action.
SEE ALSO: Doubao Proposes A New Sparse Model Architecture UltraMem
Sign up today for 5 free articles monthly!
