Yeschat LogoYeschat
Old Version

Wan 2.2: Alibaba’s Tongyi Lab Releases the World’s First Open-Source MoE Video Generation Model

In the dynamic realm of AI video generation, Wan 2.2 stands out as the world's first open-source Mixture-of-Experts (MoE) architecture video generation model, unveiled by Alibaba's Tongyi Lab on July 28, 2025. Often referred to as wan2.2 or simply wan video, this multimodal powerhouse excels in text to video (T2V), image to video (I2V), and hybrid tasks, offering cinematic-level control over lighting, composition, color grading, and complex motions like hip-hop dancing or street parkour. Fully open-sourced under an Apache 2.0 license on GitHub, Hugging Face, and ModelScope, Wan 2.2 models support up to 720p resolutions at 24fps, running efficiently on consumer-grade GPUs like the RTX 4090, making it ideal for digital art, advertising, film previsualization, and game development.The wan 2.2 models family features three variants: Wan2.2-T2V-A14B (14B parameters for superior T2V with MoE-driven layout and detail refinement), Wan2.2-I2V-A14B (14B for stable I2V synthesis reducing artifacts in stylized scenes), and Wan2.2-TI2V-5B (5B hybrid for fast 720p generation via 16×16×4 compression). This upgrade surpasses Wan 2.1 in motion fidelity—achieving reliable camera movements like pan left/right, dolly in/out, and orbital arcs—and benchmarks like Wan-Bench 2.0, where it tops competitors in semantics and aesthetics. With WanBox for all-in-one creation and editing, Wan 2.2 embodies "All in Wan, Create Anything," inviting global innovation in open video AI.

Key Features of Wan 2.2 – Next-Gen Open-Source AI Video Generation

Scalable AI Video Generation with Wan 2.2’s Mixture-of-Experts Architecture

Wan 2.2 is the world’s first open-source AI video generation model utilizing a Mixture-of-Experts (MoE) diffusion framework. By delegating denoising steps to specialized expert modules, it scales capacity without increasing computational overhead—enabling sharper frames, richer motion details, and superior temporal consistency. Compared to traditional dense diffusion models, this breakthrough delivers significantly more cinematic and coherent results in both text to video and image to video pipelines.

Cinematic Aesthetic Control in Wan 2.2 for Professional-Grade Visuals

Wan 2.2 brings cinematic-level aesthetic control to open-source AI video generation. Through prompt-based manipulation of lighting, camera movement, composition, and color grading, creators can craft compelling visual styles—from moody cyberpunk markets to serene, pastel-toned landscapes.

Unified Multi-Modal Video Creation with Wan2.2-T2V-A14B, I2V-A14B, and TI2V-5B

Wan 2.2 supports a complete range of input modalities for AI video generation. The Wan2.2-T2V-A14B model converts natural language into vivid 5-second cinematic clips at up to 720P, with impressive semantic precision and motion complexity. For static imagery, the Wan2.2-I2V-A14B model transforms images into fluid video, preserving style and spatial coherence. Need flexibility? The Wan2.2-TI2V-5B hybrid model handles both text-to-video and image-to-video tasks in a single lightweight package—capable of 720P@24fps on a single consumer GPU like the RTX 4090, making it ideal for local workflows via ComfyUI.

Fully Open-Source Wan 2.2 Models with ComfyUI Workflow Support

The entire Wan 2.2 model suite—text to video, image to video, and hybrid—is openly released and accessible via Hugging Face, GitHub, and ModelScope. With seamless ComfyUI integration, users can design node-based workflows, edit clips via timeline tools, and batch-generate assets—all within a local or cloud setup. Wan2.2’s open-source nature empowers creators, researchers, and developers to build and innovate freely within the evolving landscape of AI video generation.

Wan2.2 Model Variants: T2V, I2V, and TI2V for Text, Image, and Hybrid Video Generation

Wan2.2-T2V-A14B:High-Fidelity Text-to-Video Generation with Cinematic Precision

Wan2.2-T2V-A14B is a 14-billion-parameter text-to-video model built on the Mixture-of-Experts (MoE) architecture, offering unparalleled semantic accuracy and cinematic style control. It enables the generation of 5-second video clips at 480P and 720P, delivering visually coherent, motion-rich content directly from natural language prompts. With finely tuned capabilities for camera motion, aesthetic grading, and temporal structure, Wan2.2-T2V-A14B surpasses many leading commercial alternatives on benchmark tasks like Wan-Bench 2.0. This model is ideal for creative storytelling, advertising, and AI video research where narrative fidelity and visual polish are paramount.

Wan2.2-I2V-A14B:Stable and Stylized Image-to-Video Generation at 720P

Optimized for transforming static images into dynamic video content, Wan2.2-I2V-A14B brings cinematic expressiveness to image-to-video pipelines. Also leveraging the MoE architecture with 14 billion parameters, it supports 480P and 720P outputs while reducing common synthesis issues such as unnatural camera jitter or scene inconsistencies. The model maintains high fidelity to the source image while introducing fluid motion and spatial depth, making it ideal for digital art animation, fashion motion mockups, and cinematic content creation where visual stability and stylization are essential.

Wan2.2-TI2V-5B:Lightweight Hybrid Text & Image-to-Video Model for Local Deployment

Wan2.2-TI2V-5B is a 5-billion-parameter hybrid model designed for both text-to-video and image-to-video generation within a single unified architecture. Built on the advanced Wan2.2-VAE with a 16×16×4 compression ratio, it achieves real-time 720P at 24fps generation while remaining efficient enough to run on a single RTX 4090 GPU. This model offers an ideal balance of performance and accessibility—perfect for rapid prototyping, real-time previewing, and local workflows using ComfyUI. TI2V-5B is currently one of the fastest high-resolution open-source video generation models available for cross-modal synthesis.

Wan 2.2 vs Wan 2.1: What’s New in Next-Gen Open-Source Video AI

FeatureWan 2.1Wan 2.2
Core ArchitectureDense diffusionMixture-of-Experts (MoE) diffusion with expert hand-off across timesteps
Model VariantsT2V (14B), I2V (14B)T2V (14B), I2V (14B), TI2V Hybrid (5B)
Training DataBaseline dataset+65.6% more images, +83.2% more videos – richer motion and semantics
Aesthetic ControlBasic tagsCinematic-level labels for lighting, color, composition
Motion GenerationModerate, less controllableHigh-complexity motion, improved camera logic (tilt, orbit, dolly, etc.)
Prompt ComplianceLimited accuracyStrong prompt adherence with precise scene, motion & object control
Resolution & Frame RateUp to 720P (T2V/I2V), lower FPS720P@24fps even on single RTX 4090 (TI2V)
Performance on Consumer HardwareLimited local feasibilityTI2V runs locally on 8GB+ GPU (e.g., RTX 4090)
Use Case FlexibilityText-to-video or image-to-video onlyUnified hybrid generation + faster iteration in ComfyUI workflows
Overall Visual QualityAcceptable for baseline contentSharper frames, fewer artifacts, cinematic output polish

How to Set Up and Use Wan2.2 for AI Video Generation

  • Option 1: Local Deployment of Wan 2.2

    Wan 2.2 can be deployed locally by obtaining the official codebase and model weights from GitHub, Hugging Face, or ModelScope. These sources provide everything needed to run text-to-video, image-to-video, or hybrid generation workflows in your own environment. Once set up, you can generate 720p cinematic video content using command-line tools or integrate with ComfyUI for a visual editing experience.

  • Option 2: Use Wan 2.2 Online via the Official Web Interface

    If you prefer not to install anything, you can try Wan 2.2 directly online through Wan.Video—the official browser-based platform for fast, high-quality AI video creation. Simply enter a text or image prompt and receive a cinematic video clip in seconds, with no GPU or technical setup required.
    This option is ideal for creators, designers, and researchers looking to quickly prototype, test prompts, or generate visual concepts on the go.

    • 4 Professional Tips for Creating High-Quality Video Content with Wan 2.2

    Write Visually Descriptive and Intentional Prompts

    The key to unlocking Wan 2.2’s creative potential lies in how you write your prompts. Avoid vague instructions like “make a cool video,” and instead describe the visual elements, pacing, and emotional tone. For example, a strong prompt would be: “Create a high-energy fashion montage with fast cuts, bold text overlays, and electronic music.” The more visually specific and emotionally guided your prompt is, the more aligned the generated content will be with your creative intent.

    Use Prompt Structures That Combine Scene, Style, and Emotion

    A reliable way to guide the AI is to use structured prompts that combine three core elements: [Scene] + [Style] + [Emotion]. For instance: “Close-up shots of raindrops on glass + cinematic style + melancholic mood.” This format helps the system understand not just what to show, but how to show it and why it matters emotionally. Treat your prompt like a creative brief to a professional editor—it should communicate both the content and the mood.

    Design with Rhythm: Align Visuals to Audio Cues

    To create more professional-looking videos, consider how your visuals sync with the audio. Include instructions in your prompt that define rhythm, such as “cut on beat drops,” “build up intensity with each chorus,” or “match transitions to the tempo.” Wan 2.2 can respond to these cues with rhythm-aware editing techniques, resulting in more dynamic and engaging content that feels deliberate rather than automated.

    Iterate and Refine Through Prompt Feedback Loops

    Don’t settle for the first output—treat it as a rough cut. The real strength of Wan 2.2 lies in iterative improvement. After the initial result, analyze what’s missing or off-tone, then refine your prompt accordingly. For example: “Add more contrast and slow-motion effects in emotional scenes,” or “Reduce intro length and emphasize product close-ups.” Each round of prompting acts like a feedback loop, bringing the final output closer to your creative vision with precision.

    Use Wan 2.2 in YesChat.AI: Create Cinematic AI Videos Online

    Beyond local tools like ComfyUI, Wan 2.2 is also available on YesChat.AI, an online platform for effortless, browser-based video creation. With no installation or hardware setup required, users can generate cinematic AI videos directly from text or image prompts in seconds. Ideal for rapid prototyping, creative experimentation, and mobile workflows, YesChat.AI lowers the entry barrier for creators and researchers looking to explore Wan 2.2’s capabilities in a fast, intuitive, and accessible environment.

    FAQs About Wan 2.2

    What is Wan 2.2 and how does it redefine AI video generation?

    Wan 2.2, developed by Alibaba’s Tongyi Lab, is the world’s first open-source Mixture-of-Experts (MoE) video generation model, purpose-built for AI video generation tasks such as text to video (T2V), image to video (I2V), and hybrid workflows. Compared to previous dense models, Wan 2.2 offers cinematic fidelity, smoother motion, and scalable performance, enabling 720p@24fps generation even on consumer GPUs like the RTX 4090.

    What are the main differences between the Wan 2.2 models: Wan2.2-T2V-A14B, Wan2.2-I2V-A14B, and Wan2.2-TI2V-5B?

    The Wan 2.2 models come in three targeted variants: Wan2.2-T2V-A14B (14B parameters, optimized for high-fidelity text to video generation), Wan2.2-I2V-A14B (14B parameters, designed for stylized and stable image to video synthesis), and Wan2.2-TI2V-5B (5B parameters, a lightweight hybrid model supporting both T2V and I2V tasks at 720p on a single GPU). Each is built on the MoE architecture and optimized for different creative and technical use cases.

    How does Wan2.2-T2V-A14B achieve cinematic-level text to video generation?

    Wan2.2-T2V-A14B converts natural language prompts into visually rich, motion-consistent 5-second clips at 720p using 14B MoE parameters. It supports fine-grained control over lighting, composition, camera motion, and emotional tone—making it ideal for storytelling, concept development, and previsualization in creative industries.

    What are the advantages of using Wan2.2-I2V-A14B for image to video generation?

    Wan2.2-I2V-A14B brings stability and visual coherence to image to video generation. It transforms static images into cinematic motion while preserving artistic style and spatial layout. Leveraging MoE-based denoising, it reduces flickering, jitter, and distortion—essential for applications in digital art, stylized content creation, and animated illustration.

    When should I use Wan2.2-TI2V-5B instead of the larger 14B models?

    Wan2.2-TI2V-5B is perfect for creators seeking fast, resource-efficient hybrid video generation. It handles both text to video and image to video tasks within a compressed architecture (16×16×4 VAE), runs smoothly at 720p on a single RTX 4090, and is well-suited for real-time preview, local prototyping, and ComfyUI-based workflows without sacrificing output quality.

    What makes Wan 2.2 unique among AI video generation models today?

    Wan 2.2 is the first open-source model to combine MoE architecture with multimodal video generation (T2V, I2V, and hybrid). Its cinematic-level control, open Apache 2.0 licensing, 720p support, and real-time performance on consumer hardware make wan2.2 a uniquely accessible and powerful tool for professionals in film, advertising, gaming, and digital design.

    How can I use wan 2.2 with ComfyUI for local video generation workflows?

    Wan 2.2 offers full integration with ComfyUI, allowing users to create node-based pipelines for text to video, image to video, or hybrid tasks. After downloading the appropriate Wan 2.2 models, users can launch pre-built workflows (e.g., for Wan2.2-T2V-A14B or Wan2.2-TI2V-5B) and run local video synthesis at 720p within a visual interface—ideal for non-coders, artists, and fast iteration.

    Where can I download Wan 2.2 models and contribute to the open-source project?

    The entire wan 2.2 models suite is open-source under the Apache 2.0 license and available on GitHub, Hugging Face, and ModelScope. Users can clone the repositories, download safetensors for Wan2.2-T2V-A14B, Wan2.2-I2V-A14B, or Wan2.2-TI2V-5B, and run them locally via CLI or ComfyUI. Community contributions are encouraged through GitHub issues and pull requests—enabling global innovation in wan video creation and research.