SkyReels V2: Open Source AI Tool for Generating Unlimited Length Videos

General Introduction

SkyReels-V2 is an open source video generation model developed by SkyworkAI. It supports the generation of videos of unlimited length through advanced Diffusion Forcing techniques for both text-to-video (T2V) and image-to-video (I2V) tasks. Users can generate high-quality, movie-quality video content using text descriptions or input images. The model has a strong track record in the open source community, with performance comparable to commercial models such as Kling and Runway-Gen4. It provides flexible inference patterns suitable for developers, creators, and researchers, and the code and model weights of SkyReels-V2 are publicly available on GitHub for easy download and deployment.

Function List

• Unlimited length video generation: Support for generating videos of any length, suitable for short films to full movies.
• Text to video (T2V): Generate video content that matches the description via text prompts.
• Image to video (I2V): Generate dynamic video based on the input image, maintaining the image characteristics.
• multimodal support: Combining large-scale language modeling (MLLM) and reinforcement learning to improve video generation quality.
• Story Generation: Automatically generate video storyboards that fit the narrative logic.
• camera control: Provides a director's point of view and supports customized camera angles and movements.
• Multi-subject coherence: Ensure visual consistency of multi-character videos with the SkyReels-A2 system.
• Efficient Reasoning Framework: Supports multi-GPU reasoning to optimize generation speed and resource usage.

Using Help

Installation process

SkyReels-V2 is a Python based open source project , you need to configure the environment locally or on the server . Here are the detailed installation steps:

1. clone warehouse
   Open a terminal and run the following command to get the SkyReels-V2 code:

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

2. Creating a Virtual Environment
   It is recommended that you create a virtual environment using Python 3.10.12 to avoid dependency conflicts:

   conda create -n skyreels-v2 python=3.10
   conda activate skyreels-v2
   

3. Installation of dependencies
   Install the Python libraries needed for the project and run it:

   pip install -r requirements.txt
   

4. Download model weights
   The model weights for SkyReels-V2 are hosted at Hugging Face. download them using the following command:

   pip install -U "huggingface_hub[cli]"
   huggingface-cli download Skywork/SkyReels-V2 --local-dir ./models
   

   Make sure you have enough disk space (model sizes can be tens of gigabytes).

5. hardware requirement
   • minimum configuration: Single block RTX 4090 (24GB VRAM) with FP8 support to quantitatively reduce memory requirements.
   • Recommended Configurations: Multiple GPUs (e.g., 4-8 A100s) to support efficient parallel inference.
   • At least 32GB of system memory and 100GB of disk space.

Usage

SkyReels-V2 provides two main functions: Text to Video (T2V) and Image to Video (I2V). The following is the specific operation procedure:

Text to video (T2V)

1. Preparing Cues
   Write text prompts describing the content of the video, for example:

   A serene lake surrounded by towering mountains, with swans gliding across the water.
   

   Negative cues can be added to avoid unwanted elements:

   low quality, deformation, bad composition
   

2. Run the generated script
   modifications generate_video.py parameters, set the resolution, frame rate, etc:

   python generate_video.py --model_id "Skywork/SkyReels-V2-T2V-14B-540P" --prompt "A serene lake surrounded by mountains" --num_frames 97 --fps 24 --outdir ./output
   

   • --model_id: Select the model (e.g. 540P or 720P).
   • --num_frames: Set the video frame rate (default 97).
   • --fps: Frame rate (default 24).
   • --outdir: Output video save path.
3. View Output
   The generated video will be saved in MP4 format, e.g. output/serene_lake_42_0.mp4The

Image to video (I2V)

1. Preparing the input image
   Provide a high-quality image (e.g. PNG or JPG), making sure the resolution matches the model (default 960x544).
2. Run the generated script
   exist generate_video.py Specify the image path in the

   python generate_video.py --model_id "Skywork/SkyReels-V2-I2V-14B-540P" --prompt "A warrior fighting in a forest" --image ./input_image.jpg --num_frames 97 --fps 24 --outdir ./output
   

   • --image: Enter the image path.
   • Other parameters are similar to those of the T2V.
3. Optimized settings
   • utilization --guidance_scale(Default 6.0) Adjusts the intensity of text steering.
   • utilization --inference_steps(default 30) Controls the quality of the generation, the more steps the higher the quality but the longer it takes.
   • start using --offload Optimized memory usage for low graphics memory devices.

Featured Function Operation

1. Unlimited length video
   SkyReels-V2 uses Diffusion Forcing technology to support the generation of very long videos. Run long video inference scripts:

   python inference_long_video.py --model_id "Skywork/SkyReels-V2-T2V-14B-720P" --prompt "A sci-fi movie scene" --num_frames 1000
   

   • It is recommended to generate them in segments of 97-192 frames each, and then stitch them together with post-processing tools.
2. Story Generation
   Use the Story Generation feature of the SkyReels-A2 system to enter a plot description:

   A hero’s journey through a futuristic city, facing challenges.
   

   Running:

   python story_generate.py --prompt "A hero’s journey" --output story_video.mp4
   

   The system will generate videos containing storyboards, automatically arranging scenes and shots.

3. camera control
   pass (a bill or inspection etc) --camera_angle parameter sets the lens view (e.g. "frontal" or "profile"):

   python generate_video.py --prompt "A car chase" --camera_angle "profile" --outdir ./output
   

4. Multi-subject coherence
   SkyReels-A2 supports multi-character scenes. Provides multiple reference images to run:

   python multi_subject.py --prompt "Two characters talking" --images "char1.jpg,char2.jpg" --outdir ./output
   

   Make sure the characters are visually consistent in the video.

Optimization and Debugging

• lack of memory: Enable --quant Quantification using FP8, or --offload Offloads some calculations to the CPU.
• Generating quality: Increase --inference_steps(e.g., 50) or adjust --guidance_scale(e.g. 8.0).
• Community Support: Check GitHub Issues for problems or join the SkyReels Community Discussion.

application scenario

1. Short video creation
   Creators can use the T2V feature to quickly generate short video clips from text, suitable for social media content production.
2. Movie pre-production
   Directors can utilize the unlimited length video and story generation features to create movie trailers or concepts and reduce upfront costs.
3. Virtual E-Commerce Showcase
   Use the I2V function to turn product pictures into dynamic videos to show how the product is used in a virtual scene.
4. Educational animation
   Teachers can generate instructional animations from text descriptions to visualize complex concepts, such as the process of a science experiment.
5. game development
   Developers can generate game scenes or character animations to be used as material for prototyping or transitions.

QA

1. What resolutions does SkyReels-V2 support?
   Currently supports 540P (960x544) and 720P (1280x720), with the possibility of expanding to higher resolutions in the future.
2. How much video memory do I need to run it?
   A single RTX 4090 (24GB) can run basic reasoning, and multi-GPU configurations can accelerate raw and grown video.
3. How to improve the quality of generated videos?
   Increase the number of reasoning steps (--inference_steps), optimize prompt words, or use high-quality input images.
4. Does it support real-time generation?
   Currently offline generation, real-time generation requires higher hardware support and may be optimized in the future.
5. Are model weights free?
   Yes, SkyReels-V2 is completely open source and the weights can be downloaded for free from Hugging Face.