Wan 2.2 AI Video Generator: The Four Variants Explained (2026 Guide)

The first time I set up Wan 2.2, I picked the biggest checkpoint, queued a text prompt, and waited. Four hours later I had a clip that was worse than what a browser tool gave me in three minutes, because I had loaded the image-to-video model and fed it nothing but text.

That mistake is common, and it is not really the user's fault. "Wan 2.2" is one name for five different checkpoints, each trained for a different job, all released under the same licence, all downloadable from the same place. Pick the wrong one and the model looks broken when it is simply being asked to do something it was not built for.

This guide is the map I wish I had: what each Wan 2.2 variant does, which one to load for which job, what it costs in hardware or credits, and where each one fails.

TL;DR

• Wan 2.2 is not one model. It ships as T2V-A14B, I2V-A14B, TI2V-5B, S2V-14B and Animate-14B, plus community branches — five different jobs under one version number.
• T2V-A14B is for text-only clips, I2V-A14B animates a still, and TI2V-5B does both at 720p on consumer hardware.
• S2V-14B needs an audio track and produces talking or singing video; Animate-14B needs a driving video and retargets movement onto your character.
• The 5B model is the practical entry point for local runs: 720p at 24fps, and the project documents it running on a single consumer flagship card.
• The A14B models are the quality tier, using a mixture-of-experts design that splits denoising across timesteps.
• Everything here is Apache 2.0, which is why free browser tools can host Wan 2.2 and why no one can put a paywall on the weights.

Run Wan 2.2 Free in the Browser →

What "Wan 2.2" actually covers

Wan 2.2 launched on 28 July 2025 as Alibaba's second major open video generation release, under Apache 2.0. Two variants arrived the same day (T2V-A14B and I2V-A14B), a smaller hybrid model shipped alongside them, and two more were added over the following two months.

Checkpoint	Released	Job	Resolutions	Notes
T2V-A14B	Jul 2025	Text → video	480p, 720p	MoE model, cinematic aesthetic data
I2V-A14B	Jul 2025	Image → video	480p, 720p	Same MoE design, first-frame conditioning
TI2V-5B	Jul 2025	Text or image → video	720p @ 24fps	High-compression VAE, consumer-GPU target
S2V-14B	Aug 2025	Audio + image → talking video	480p, 720p	Speech-driven, lip-sync oriented
Animate-14B	Sep 2025	Driving video → character motion	Varies	Animation and character replacement

Two implementation details explain most of the behavioural differences.

The A14B models use a mixture-of-experts architecture: instead of one network denoising every step, specialised experts handle different parts of the denoising trajectory. The stated goal is more capacity at the same compute cost, and the practical effect is better motion and lighting than Wan 2.1 at similar generation time.

The 5B model uses a high-compression VAE (16×16×4), which is the reason it can output 720p at 24fps on a single consumer card. Compression is a trade: less latent detail to preserve means faster generation and slightly softer fine texture.

Which variant for which job

This is the table I check before loading anything.

Your job	Variant	Why
Turn a written idea into a clip	T2V-A14B	Built for text conditioning
Animate a product photo	I2V-A14B	Preserves the still's composition
Iterate 20 prompt variations cheaply	TI2V-5B	Fast, 720p, low VRAM
Make a character talk or sing	S2V-14B	Audio drives the mouth and head motion
Apply a dance or gesture to a character	Animate-14B	Uses a driving video, not a prompt
Run on a laptop or a small GPU	TI2V-5B	Only variant designed for that

If you are unsure, start with TI2V-5B. It does both text and image conditioning, it is the fastest to test, and it will tell you within a few generations whether your idea is a Wan idea at all.

Test Wan 2.2 Without a GPU →

If you want output today, start here: Launch Wan 2.2 Now →

T2V-A14B: the text-to-video tier

This is the checkpoint most people mean by "Wan 2.2 AI video generator", and it is the one that benefits most from prompt structure.

What it does well: complex motion, camera moves that stay coherent, and lighting that looks chosen rather than averaged. Alibaba's release notes credit a curated aesthetic dataset with labels for lighting, composition, contrast and colour tone, which is why prompting for light direction actually does something on this model.

Where it fails: precise text rendering inside the frame, multi-character scenes that require distinct identities, and anything longer than the clip window. It also has no audio, so dialogue scenes need a second pass or a different variant.

A prompt order that works consistently:

1. Subject and wardrobe.
2. Action, in verbs, with the specific body part doing it.
3. Environment and time of day.
4. Camera: shot size, movement, lens feel.
5. Light: direction and quality.

Putting the camera before the action is the most common reason a clip ignores the action.

I2V-A14B: animating a still

Image-to-video is the variant that changes the economics of product and character work, because the still carries the identity and the model only has to move it.

What it does well: keeping the composition of the source frame, adding believable motion to a subject that is already correctly lit, and holding a product's shape. If your still is clean, the output is usually usable on the first or second try.

Where it fails: source images with soft focus, extreme angles, or ambiguous edges. The model will happily invent motion for a blurry subject and produce the classic "melting edges" artefact. Also, if your prompt describes a subject that is not in the image, the model fights the frame and produces drift.

The prompting rule: describe only what moves. Do not re-describe the image. Say what the subject does, where the camera goes, and what the light does. Everything else is already decided by the frame.

TI2V-5B: the one that fits your machine

The 5B model is the reason "run Wan locally" is realistic advice rather than a flex. The project documents it at 720p, 24fps, on a consumer-grade card of the 4090 class, and its high-compression VAE is what makes that possible.

What it does well: speed, iteration, and both text and image inputs from one checkpoint. If you are testing whether a shot works, this is the one to load, because you can run ten attempts in the time an A14B run takes for two.

Where it fails: fine texture, small faces, distant subjects, and high-frequency detail like foliage or fabric weave. It also shows compression artefacts more readily on hard edges during fast motion. Quantised community builds push the memory requirement lower still, at some quality cost.

Ready to try it yourself? Try Wan 2.2 Free →

The honest use case: 5B for exploration, A14B for the shots that survive exploration.

S2V-14B: speech-driven video

S2V was added on 26 August 2025 as an audio-driven cinematic model: you supply a portrait or character image plus an audio track, and it generates video where the mouth and head follow the speech. Speech synthesis support via CosyVoice was added shortly after, so you can go from a script to a performance without recording.

What it does well: talking-head and performance shots, singing clips, and any content where the audio is the spine. It is also the variant that most directly replaces a traditional avatar pipeline.

Where it fails: overlapping speakers, whispering, heavy accents at low audio quality, and faces turned far from camera. Lip-sync degrades with head rotation, which is why the best outputs keep the subject roughly front-facing.

Rule of thumb: clean, single-speaker, close-mic audio in; believable performance out. Compressed phone audio in; uncanny mouth out.

Animate-14B: movement from a driving video

Released on 19 September 2025, this one is different in kind: instead of a prompt describing motion, you give it a driving video and it applies that movement to your character, for either animation or character replacement.

What it does well: dance choreography, gesture transfer, and re-performing a reference clip with your own character. Because the motion is measured rather than imagined, the output is more faithful to a specific movement than any prompt could produce.

Where it fails: heavy occlusion in the driving video, very different body proportions between driver and character, and fast camera movement in the source. It also needs a usable driving clip, which is a production step in itself. For a closer look at how it stacks up against other models, see Kling 2.6 Motion Control vs Wan 2.2 Animate. If you want to test it without installing anything, the Wan 2.2 Animate tool works in the browser.

Local versus browser: the real trade

	Local (ComfyUI or CLI)	Free browser tool
Cost	GPU time and electricity	Zero per clip
Setup	Environment, weights, 20–100 GB of downloads	None
Control	Every sampler and scheduler exposed	Prompts and a few settings
Filters	None	Platform policy applies
Queue	Yours to manage	Shared
Ceiling	Best quality if you tune it	Good enough for most social work

Both routes run the same weights. The difference is who owns the friction. For a first test, the browser route wins on time. For volume or for anything the platform filters, local wins on control.

Skip the Install: Generate in the Browser →

Cost, computed rather than guessed

Want to see the difference on your own footage? Start creating with Wan 2.2 →

Wan 2.2 is free to run; the hardware is not. A rough model for local work: a 5-second clip on a consumer card at a few hundred watts, running for several minutes, costs cents in electricity. The same clip through a hosted endpoint priced at $0.10 per second costs $0.50. Twenty iterations is $10 hosted, or a couple of dollars of power locally, ignoring the card you already own.

That math changes at very low volume. If you need four clips a month, a free browser tool is cheaper than any of it, because your time matters more than the marginal cost.

Failure modes by variant

Symptom	Likely variant problem	Fix
Motion ignores the prompt	T2V, camera described before action	Reorder: subject, action, environment, camera
Edges melt and smear	I2V, blurred or ambiguous source	Sharpen the still, simplify the background
Faces are mushy	5B at 720p, subject too small	Reframe closer, or move to A14B
Mouth is out of sync	S2V, head turning or noisy audio	Use close-mic audio, keep the face to camera
Character distorts	Animate, proportion mismatch	Match framing and body scale to the driver
Output drifts from the still	Prompt re-describes the image	Describe only motion

The Bottom Line

Choosing a Wan 2.2 "generator" is really choosing between five checkpoints, and the version number tells you nothing about which job a checkpoint was trained for. Text-to-video for ideas, image-to-video for stills, 5B for iteration, S2V for speech, Animate for driving video.

Start with the job, pick the variant, and only then decide whether to run it locally or in a browser. If you want to see what the family does before downloading anything, generate a Wan 2.2 clip for free and use that as your baseline.

FAQ

What is Wan 2.2 AI video generator?

Wan 2.2 is Alibaba's open-weight video generation model family, released under Apache 2.0 in July 2025. It includes text-to-video, image-to-video, a compact hybrid model, a speech-driven model and a character animation model, all downloadable and all runnable locally or through browser tools.

Which Wan 2.2 model should I use?

For text-only clips use T2V-A14B; for animating a still use I2V-A14B; for speed and lower memory use TI2V-5B; for talking or singing video use S2V-14B; for transferring movement from a reference clip use Animate-14B. If you are unsure, start with TI2V-5B because it handles both text and image inputs.

Can Wan 2.2 run on a normal GPU?

The TI2V-5B model is the one designed for that: the project documents 720p at 24fps on a consumer-grade card of the 4090 class, and quantised community builds bring the requirement down further. The A14B models are heavier and generally expect a higher-memory card or a hosted service.

Is Wan 2.2 free for commercial use?

Yes. Wan 2.2 is released under Apache 2.0, which permits commercial use of the weights. Platform terms still apply to the service you use to run it, and any third-party model you add to a workflow carries its own licence.

Does Wan 2.2 generate audio?

Not in the T2V, I2V or TI2V checkpoints, which are silent. Audio enters through S2V-14B, which is driven by a supplied audio track and can use CosyVoice for speech synthesis. Later hosted Wan versions add audio generation, but that is a different product from the open 2.2 checkpoints.

What is the difference between Wan 2.2 and Wan 2.2 Animate?

Animate-14B is a separate checkpoint released in September 2025 that takes a driving video and applies its movement to your character, for animation or replacement. The base 2.2 models generate motion from text or from a still, not from a reference performance.

References

• Wan2.2 official repository: variant list, MoE architecture, high-compression VAE for TI2V-5B, 720p 24fps note, Apache 2.0 licence, ComfyUI and Diffusers integration.
• Wan2.2-T2V-A14B on Hugging Face: text-to-video checkpoint, 480p and 720p support.
• Wan2.2-TI2V-5B on Hugging Face: compact hybrid checkpoint and its compression approach.
• Wan2.2-S2V-14B on Hugging Face: speech-driven video model.
• ComfyUI Wan 2.2 tutorial: node setup for local generation.
• Wan official site: hosted playground for the Wan family.

Related guides

• Wan video models compared — where Wan 2.2 sits in the wider family, open weights versus API.
• Wan 2.2 free: how to use it online — the free tool routes and prompt patterns.
• Wan 2.2 Animate free guide — the driving-video workflow in practice.
• Wan 2.2 generation guide — settings and prompting detail per model.
• Wan 2.2 free unlimited — what "unlimited" actually covers.

Start With Wan 2.2 Free →

Need Longer Clips? Try Wan 2.7 on Pollo →