Introduction to Wan 2.1 Models

Introduction

Wan 2.1 represents a groundbreaking suite of open-source video foundation models that sets new standards in video generation technology. This article explores its key features and capabilities.

There is a new leader in open source video generation! Alibaba's new Wan 2.1 model is now the leading open weights model in the Artificial Analysis Video Arena, surpassing former titleholder Mochi 1

Wan 2.1 is a 14B parameter model (1.3B variant also released) and stands out for its ability to generate realistic looking video with high-fidelity motion.

Key details regarding Wan 2.1: ➤ The 14B model is available in image to video, and text to video variants. The 1.3B model only supports text to video ➤ The 14B parameter model supports 720p output while the 1.3B model outputs at 480p ➤ Generates natively at 16 fps. Compared to other models that generate at 24 fps, this can result in a slight stuttering effect ➤ Supports multilingual text input in both English and Chinese ➤ The 1.3B model only requires 8.2GB of VRAM, allowing many consumer grade GPUs to support inferencing the model. Alibaba claims a RTX 4090 can generate a 5 second 480p video in ~4 minutes

See thread below for comparisons between Wan 2.1, Veo 2 and other leading models in our arena 🧵

Key Features

State-of-the-Art Performance

Wan 2.1 consistently outperforms both existing open-source models and commercial solutions across multiple benchmarks. Its comprehensive evaluation across 14 major dimensions and 26 sub-dimensions demonstrates superior capabilities in motion quality, visual quality, style rendering, and multi-targeting scenarios.

Consumer-Friendly Hardware Requirements

One of the most remarkable aspects of Wan 2.1 is its accessibility. The T2V-1.3B model requires only 8.19 GB VRAM, making it compatible with consumer-grade GPUs. On an RTX 4090, it can generate a 5-second 480P video in approximately 4 minutes without any optimization techniques.

Multi-Task Capabilities

Wan 2.1 excels in multiple tasks including:

• Text-to-Video generation
• Image-to-Video conversion
• Video Editing
• Text-to-Image generation
• Video-to-Audio processing

Advanced Text Generation

A unique feature of Wan 2.1 is its ability to generate both Chinese and English text within videos, making it the first video model with bilingual text generation capabilities.

Model Variants

Wan2.1-I2V-14B

• Supports both 480P and 720P resolution
• Outperforms leading closed-source models
• Excels in generating complex visual scenes and motion patterns
• Takes both text and images as input

Wan2.1-T2V-14B

• Supports 480P and 720P resolution
• Sets new SOTA performance benchmarks
• Features bilingual text generation
• Demonstrates superior motion dynamics

Wan2.1-T2V-1.3B

• Optimized for consumer GPUs
• Requires only 8.19 GB VRAM
• Generates 480P videos
• Achieves performance comparable to some closed-source models through pre-training and distillation

Powerful Video VAE

The Wan-VAE component delivers exceptional efficiency in:

• Encoding and decoding 1080P videos of any length
• Preserving temporal information
• Providing a robust foundation for video and image generation

Conclusion

Wan 2.1 represents a significant advancement in video generation technology, offering state-of-the-art performance while maintaining accessibility for consumer-grade hardware. Its comprehensive feature set and multiple model variants make it a versatile solution for various video generation needs.