SkyReels-V2: Infinite-Length Film Generative Model by Skywork
SkyReels-V2 is the first open-source infinite-length film generative model using Diffusion Forcing. Built on Wan2.1, it supports T2V, I2V, and autoregressive video generation.
SkyReels-V2
VideoText-to-VideoImage-to-VideoDiffusion ForcingInfinite-LengthOpen SourceSkyReels-V2 is an infinite-length film generative model by Kunlun Tech (Skywork), employing an AutoRegressive Diffusion-Forcing architecture for SOTA open-source video generation. Built on Wan2.1, it supports Text-to-Video (T2V), Image-to-Video (I2V), and infinite-length Diffusion Forcing (DF) generation via both synchronous and asynchronous inference modes. Available in 14B and 1.3B parameter sizes at 540P and 720P resolutions. LoRA adapters are also available for fine-tuning. ComfyUI support via Kijai's WanVideoWrapper.
| Developer | Kunlun Tech (Skywork AI) |
| Release Date | 2025-04-21 |
| Architecture | AutoRegressive Diffusion-Forcing Transformer |
| Base Model | Wan2.1 (Wan-AI) |
| License | SkyReels-V2 Community License |
| Model Sizes | 14B (full) / 1.3B (lightweight) |
| Capabilities | Text-to-Video (14B), Image-to-Video (14B), Diffusion Forcing (14B/1.3B) : infinite-length autoregressive |
| Output Resolutions | 540P (960×544) / 720P (1280×720) |
| Max Frames | 97~1457 frames (4s–60s+ at 24fps) |
| VRAM (540P) | ~51.2 GB (14B) / ~14.7 GB (1.3B) |
Overview
SkyReels-V2 is the first open-source video generative model employing an AutoRegressive Diffusion-Forcing architecture, achieving SOTA performance among publicly available models. It synergizes Multi-modal Large Language Models (MLLM), Multi-stage Pretraining, Reinforcement Learning, and the Diffusion Forcing Framework to enable infinite-length film generation.
Model Variants
Three model types are available, each with multiple resolutions and precision options:
| Type | Description | Available Sizes | Precision |
|---|---|---|---|
| Diffusion Forcing (DF) | Infinite-length autoregressive video generation; supports both T2V and I2V tasks in synchronous or asynchronous modes | 14B (540P/720P), 1.3B (540P) | fp16, fp8_e4m3fn, fp8_e5m2 |
| Text-to-Video (T2V) | Standard text-to-video generation | 14B (540P/720P) | fp16, fp8_e4m3fn, fp8_e5m2 |
| Image-to-Video (I2V) | Image-conditioned video generation | 14B (540P/720P) | fp16, fp8_e4m3fn, fp8_e5m2 |
| SkyReels A2 | Controllable video generation for assembling arbitrary visual elements; Apache-2.0 | 14B (fp8) | fp8_e4m3fn |
Key Features
- Infinite-Length Generation: Uses Diffusion Forcing with autoregressive frame prediction to generate arbitrarily long videos beyond single-clip limits
- Synchronous & Asynchronous Inference: Synchronous for faster generation, asynchronous for better instruction following and visual consistency
- Video Extension: Extend existing videos with the DF model by passing a video path
- Start/End Frame Control: Control both first and last frames of generated videos
- Prompt Enhancement: Built-in prompt enhancer using Qwen2.5-32B-Instruct
- TeaCache Acceleration: Optional caching for faster inference
- Offload Support: CPU offloading reduces peak VRAM requirements
ComfyUI Setup
These models are packaged by Kijai for use with the WanVideoWrapper custom node. Place diffusion models in ComfyUI/models/diffusion_models/ and LoRAs in ComfyUI/models/loras/.