SkyReels-V2: Infinite-Length Video Generation by Skywork
SkyReels-V2 is Skywork's open Diffusion Forcing video model for infinite-length generation, with 14B and 1.3B checkpoints at 540P and 720P.
SkyReels-V2
VideoText-to-VideoImage-to-VideoDiffusion ForcingInfinite-LengthSkywork AI's open video model built on an AutoRegressive Diffusion-Forcing architecture. It generates arbitrarily long clips beyond single-clip limits instead of treating each shot in isolation, and ships as 14B and 1.3B checkpoints at 540P and 720P in Text-to-Video, Image-to-Video and Diffusion Forcing builds, plus LoRA adapters for fine-tuning.
| Developer | Kunlun Tech (Skywork AI) |
| Release Date | 2025-04 |
| Architecture | AutoRegressive Diffusion-Forcing Transformer |
| Base Model | Wan2.1 (Wan-AI) |
| License | SkyReels-V2 Community License |
| Model Sizes | 14B (540P / 720P), 1.3B (540P) |
| Capabilities | Text-to-video, image-to-video, infinite-length Diffusion Forcing, video extension, start/end frame control |
| Precision | fp16, fp8_e4m3fn, fp8_e5m2 |
What is SkyReels-V2?
SkyReels-V2 is the first open-source video generative model built on an AutoRegressive Diffusion-Forcing architecture, combining a multimodal LLM, multi-stage pretraining, reinforcement learning and the Diffusion Forcing framework. Diffusion Forcing predicts frames autoregressively rather than generating one fixed-length clip, which is what allows infinite-length film generation: a video can keep going past the limits of a single window, and an existing video can be extended by feeding it back in.
Three build types ship at 540P and 720P: standard Text-to-Video, Image-to-Video, and the Diffusion Forcing models that support both tasks in synchronous (faster) or asynchronous (better instruction following and consistency) inference. A 1.3B DF checkpoint targets consumer GPUs, and LoRA adapters cover 720P T2V and I2V fine-tuning.
Guides and workflows related to this model series.
Comments
Sign in with GitHub to join the discussion.