Zen Director is a 5B parameter text-to-video and image-to-video generation model. Based on Wan 2.2, it creates high-quality videos from text descriptions and images with controllable motion.

Built on Wan-AI/Wan2.2-TI2V-5B - Text-to-Image-to-Video model with 5B parameters.

Note: This is based on Wan 2.2. Wan 2.5 is announced but not yet open-source. We will upgrade to Wan 2.5 when it becomes available.

• Text-to-Video: Generate videos from text descriptions
• Image-to-Video: Animate static images into videos
• High Resolution: Supports high-quality video generation
• Efficient: Optimized MoE architecture for fast inference

• Architecture: Mixture-of-Experts (MoE) Transformer
• Parameters: 5B total
• Base: Wan 2.2 TI2V
• Resolution: Up to 1280x720
• Frame Rate: 24 FPS
• Duration: Up to 5 seconds

pip install diffusers transformers accelerate torch
pip install av opencv-python pillow

from diffusers import DiffusionPipeline
import torch

# Load the model
pipe = DiffusionPipeline.from_pretrained(
    "zenlm/zen-director",
    torch_dtype=torch.float16
)
pipe = pipe.to("cuda")

# Generate video from text
prompt = "A serene sunset over a calm ocean with waves gently lapping at the shore"
video = pipe(prompt, num_frames=120, height=720, width=1280).frames

# Save video
from diffusers.utils import export_to_video
export_to_video(video, "output.mp4", fps=24)

from PIL import Image

# Load starting image
image = Image.open("input.jpg")

# Generate video from image
video = pipe(
    prompt="Animate this image with gentle camera movement",
    image=image,
    num_frames=120
).frames

export_to_video(video, "animated.mp4", fps=24)

• Inference Speed: ~2-3 seconds/frame on A100
• Memory: Requires 24GB+ VRAM for full resolution
• Quantization: FP16 recommended for consumer GPUs

• ✅ v1.0 - Wan 2.2 TI2V-5B base (current)
• 🔄 v2.0 - Upgrade to Wan 2.5 when open-source
• 📋 Future - Fine-tuning for specific styles and domains

• Requires high-end GPU (24GB+ VRAM recommended)
• Video duration limited to 5 seconds
• Best results with detailed, specific prompts
• Some motion artifacts in complex scenes

• GitHub: https://github.com/zenlm
• Zen Gym (Training): https://github.com/zenlm/zen-gym
• Zen Engine (Inference): https://github.com/zenlm/zen-engine

@misc{zen-director-2025,
  title={Zen Director 5B: Text-to-Video Generation Model},
  author={Zen AI Team},
  year={2025},
  howpublished={\url{https://huggingface.co/zenlm/zen-director-5b}}
}

@article{wan2024,
  title={Wan 2.2: High-Quality Video Generation},
  author={Wan-AI Team},
  journal={arXiv preprint},
  year={2024}
}

Apache 2.0

Note: Based on Wan 2.2. Will be upgraded to Wan 2.5 when it becomes open-source.

Part of the Zen AI ecosystem.

zen-director is based on CogVideoX-5B

We are grateful to the original authors for their excellent work and open-source contributions.

• Repository: https://huggingface.co/THUDM/CogVideoX-5B
• Base Model: CogVideoX-5B
• License: See original repository for license details

• Adapted for Zen AI ecosystem
• Fine-tuned for specific use cases
• Added training and inference scripts
• Integrated with Zen Gym and Zen Engine
• Enhanced documentation and examples

If you use this model, please cite both the original work and Zen LM:

@misc{zenlm2025zen-director,
    title={Zen LM: zen-director},
    author={Hanzo AI and Zoo Labs Foundation},
    year={2025},
    publisher={HuggingFace},
    howpublished={\url{https://huggingface.co/zenlm/zen-director}}
}

Please also cite the original upstream work - see https://huggingface.co/THUDM/CogVideoX-5B for citation details.