Wan 2.2: A Mixture-of-Experts Model for Video Generation

Wan 2.2 is the latest and greatest in open-source AI video generation, developed by Alibaba's Tongyi Lab. This model introduces new architectural innovations that significantly advance the field of text-to-video and image-to-video generation while maintaining computational efficiency, making it affordable for many use cases. This is the A14B model that can output 480p and 720p videos. There is also a smaller 5B model that is consumer GPU friendly.

About the tech: Mixture-of-Experts (MoE) Architecture

The model leverages Mixture of Experts or MoE architecture and uses 2 expert models under the hood for the diffusion denoising process.

• •High-noise expert: Processes early denoising stages, focusing on overall layout and structure
• •Low-noise expert: Manages final stages, refining video details and reducing artifacts

This two model approach results in 14 billion active parameters per inference step and a total of 27 billion over all parameters combining both the models. The transition between the experts is decided based on the signal to noise ration (SNR) helping the pipeline intelligently transition between the two experts without sacrificing output quality

Wan 2.2 also brings substantial improvements over its predecessor through expanded training data, featuring 65% more images and 80% more videos. This enables advance motion generations helping generate complex body movements, dynamic scene transitions and fluid camera controls. It can also simulate realistic physics scenarios and object interactions. This makes it effective for character animations, sport scenes and other cimenatic sequences.

Another big leap due the added training data is tight control over lighting, composition, contrast and color tone. Wan 2.2 offers over 60 controllable parameters that enale control for camera aware prompting like "aerial orbit," "handheld tracking shot," or specific lighting requirements.

Use cases

Just like other Text to Video and Image to video models this model can be used for a range of use cases. Use to generate cinematic visuals for a project you are working on or generate short social media ads with a product in focus. You can also use the model to create simple animations that can be used as a website background or on a slide deck. The lower costs compared to a lot of open models out there makes it the first choice before trying other video generator model.

License details

Wan 2.2's open-source nature under the Apache 2.0 license makes it the best choice for a range of use cases including commercial and educational purposes.

Effective Guide to Wan2.2 TI2V-5B (Text-to-Video AI)

Wan2.2 TI2V-5B is a Mixture-of-Experts text-to-video model that converts prompts or images into cinematic 720p clips at 24 fps on a single consumer GPU. Follow this guide to fine-tune parameters for different goals and streamline your workflow.

Key Parameters

• •prompt (string, required): vivid, descriptive text.
• •image (string, required for image-based clips): high-res JPEG/PNG URL.
• •resolution (enum: “480p”/“720p”): 720p for production; 480p for rapid drafts.
• •aspect_ratio (enum: “16:9”/“9:16”): landscape or portrait framing.
• •num_frames (int: 81–100): clip length and smoothness.
• •frames_per_second (int: 5–24): motion fluidity.
• •go_fast (bool): true for speed; false for maximum quality.
• •sample_shift (float: 1–20): stylistic variation control.
• •seed (int): fixed seed for reproducibility.

Use-Case Presets

1. Marketing & Advertising

Generate eye-catching 5–8 s teasers.

• •resolution: “720p”
• •aspect_ratio: “16:9” or “9:16” for social ads
• •num_frames: 95 (≈6 s at 16 fps)
• •frames_per_second: 16
• •go_fast: true
• •sample_shift: 8
• •seed: omit or random

2. Film Pre-Visualization

Prototype camera moves and lighting.

• •resolution: “480p”
• •aspect_ratio: “16:9”
• •num_frames: 100 (≈5 s at 20 fps)
• •frames_per_second: 20
• •go_fast: false
• •sample_shift: 3
• •seed: specify for exact repeats

3. E-Learning & Explainers

Craft clear narrative sequences.

• •resolution: “720p”
• •aspect_ratio: “16:9”
• •num_frames: 81 (≈5 s at 16 fps)
• •frames_per_second: 16
• •go_fast: false
• •sample_shift: 5

4. Game Cinematics & Trailers

Visualize concept animations.

• •resolution: “720p”
• •aspect_ratio: “16:9”
• •num_frames: 100
• •frames_per_second: 24
• •go_fast: false
• •sample_shift: 15
• •seed: lock for consistency

5. Social Content Creation

Short, engaging loops for mobile.

• •resolution: “480p”
• •aspect_ratio: “9:16”
• •num_frames: 81
• •frames_per_second: 12
• •go_fast: true

Pro Tips

1. •Be specific: “golden hour sunlight” or “wind-blown leaves.”
2. •Anchor with a detailed image URL for realistic motion.
3. •Lower fps (≤12) for stylized, stutter-free loops.
4. •Toggle go_fast off for richer textures.
5. •Use seed to reproduce key concept shots.
6. •Adjust sample_shift to explore subtle style shifts.

Harness these presets and tips to unlock Wan2.2 TI2V-5B’s full potential!

FAQs