Wan2.2: Large-scale video generative models
Open-source diffusion framework for multi-modal video generation.
Learn more about Wan2.2
Wan2.2 is a collection of diffusion-based video generation models designed for various input modalities and output specifications. The architecture incorporates a Mixture-of-Experts approach to separate denoising across timesteps using specialized expert models, and uses a custom VAE with 16x16x4 compression ratios. The framework supports inference on consumer-grade GPUs and has been integrated into Diffusers, ComfyUI, and ModelScope. Common deployments include text-to-video generation, image-to-video synthesis, character animation, and audio-driven video creation.
Mixture-of-Experts architecture
Uses specialized expert models for different denoising timesteps in the diffusion process, increasing model capacity without proportional increases in computational cost during inference.
Multiple model variants
Offers 5B and 14B parameter models supporting different generation tasks (T2V, I2V, TI2V, character animation, audio-driven) with 720P output at 24fps, runnable on consumer GPUs like RTX 4090.
Aesthetic control and motion complexity
Trained on expanded datasets with curated aesthetic labels for lighting, composition, and color tone, plus significantly increased video and image data to improve motion generation and semantic understanding across diverse scenarios.
from wan22.pipeline import Wan2Pipeline
pipeline = Wan2Pipeline.from_pretrained(
"Comfy-Org/wan2.2_ti2v_5B_fp16",
vae="wan2.2_vae",
text_encoder="umt5_xxl"
)
video = pipeline(
prompt="A cinematic shot of waves crashing on a beach at sunset",
num_frames=81,
height=480,
width=832
)
video.save("output.mp4")Related Repositories
Discover similar tools and frameworks used by developers
yolov5
Real-time object detection with cross-platform deployment support.
openvino
Convert and deploy deep learning models across Intel hardware.
CodeFormer
Transformer-based face restoration using vector-quantized codebook lookup.
NAFNet
Efficient PyTorch architecture for image restoration tasks.
streamlit
Python framework for reactive data web applications.