SkyworkAI /SkyReels-V2

About SkyReels-V2

SkyReels-V2 is an open video model you run on your own machine. It makes long clips, and can keep going for endless ones. You get text-to-video, image-to-video, clip extension and start/end frame control. The 14B model needs 51.2 GB of VRAM, and the license is unclear.

Best for: Technical makers who want a long-form video model running on their own GPU.

• Makes long videos, and can keep generating to make them endless
• Handles text-to-video, image-to-video, clip extension and start/end frames
• Runs on one GPU or several, with sync and async modes
• Ships diffusers pipelines for T2V and I2V in Python
• Includes SkyCaptioner-V1 for video captions and a prompt enhancer
• Saves memory with offloading, TeaCache and a smaller base frame count

1. 1
   Clone the repo and install the requirements: the command below, cd SkyReels-V2, then the next command (tested on Python 3.10.12)

   git clone https://github.com/SkyworkAI/SkyReels-V2

   Copy

   pip install -r requirements.txt

   Copy
2. 2
   Download the model weights you want from Hugging Face or ModelScope, then set the model path
3. 3
   Generate a long video: the command below

   python3 generate_video_df.py --model_id Skywork/SkyReels-V2-DF-14B-540P --resolution 540P --ar_step 0 --base_num_frames 97 --num_frames 257 --overlap_history 17 --prompt "..." --addnoise_condition 20 --offload

   Copy
4. 4
   For image-to-video add --image ${image_path}; to extend a clip add --video_path ${video_path}
5. 5
   Or use the diffusers pipelines: SkyReelsV2DiffusionForcingPipeline for T2V and SkyReelsV2DiffusionForcingImageToVideoPipeline for I2V