SkyReels-V3

Skywork's third SkyReels generation splits video work across three open-weight pipelines: a 14B reference-to-video model that preserves the identity of 1 to 4 people or objects from supplied images, a 14B video extension model that continues existing footage with cinematographic transitions, and a 19B audio-driven talking avatar model, all generating at 720p. Inference code landed on GitHub in January 2026 with checkpoints auto-downloaded from Hugging Face or ModelScope; the stack wants Python 3.12 and CUDA 12.8, and cards under 24 GB can still run it by enabling FP8 weight-only quantization with block offload or dropping output to 540p or 480p. The weights ship under the Skywork Community License, which permits commercial use, and the same models back Skywork's paid API for teams that skip local hosting. Multi-subject consistency is the headline capability: reference conditioning makes this one of the few open options for keeping specific faces and products stable across generated shots, a task where prompt-only video models routinely drift.