Wan 2.2 Online: How to Run It in a Browser in 2026 (Free vs Paid vs Local)

Introduction

Wan 2.2 is a year old, superseded twice, and still the most practical open video model for anyone who does not want to rent a cloud GPU at 3am. Its weights are Apache 2.0. Its 5B variant generates 720p at 24fps on a single RTX 4090. And its ecosystem — the LoRAs, the quantised builds, the ComfyUI graphs — is the largest of any Wan release, because it shipped at the moment open video generation actually worked.

Which is exactly why "wan 2.2 online" is still a live search. People are not looking for a flagship. They are looking for a way to run a specific checkpoint without installing anything, and the information they find is usually a landing page pretending Wan 2.2 is the newest model on earth.

This guide is the practical version: the three online routes that actually exist, what each one costs in money and waiting, which of the four Wan 2.2 checkpoints you need, the settings that decide whether the output is usable, and the honest answer to when you should stop using Wan 2.2 entirely.

TL;DR

• There are three online routes, and they are not interchangeable: a hosted browser tool (zero setup, capped control), serverless inference APIs (pay per call, scriptable), and Hugging Face Spaces (free GPU queues, variable quality).
• Know which checkpoint you want before you pick a route. T2V-A14B and I2V-A14B are 480p/720p MoE models needing ~80GB for single-GPU inference; TI2V-5B is the 720p@24fps model that runs on 24GB.
• TI2V-5B is the one that makes "online" sensible. A 16×16×4 VAE compression keeps it small enough for a 4090 and fast enough for a browser tool to feel interactive.
• Wan 2.2 base checkpoints are silent. T2V-A14B, I2V-A14B and TI2V-5B output no audio track at all. If your deliverable needs sound, that is a 2.5+ feature, not a 2.2 setting.
• 720p is the ceiling, not a starting point. The 720p TI2V task uses 1280×704; image-to-video inherits the source image's aspect ratio.
• Switch to Wan 2.7 when you need audio, 1080p, multi-shot, or 15-second clips. Stay on 2.2 when you need free, fast, LoRA-compatible, and good enough.

Run Wan 2.2 in Your Browser

What "Online" Actually Means for Wan 2.2

The word covers four genuinely different setups, and picking the wrong one is why people conclude the model is bad:

Route	Setup time	Who runs the GPU	Control	Best for
Hosted browser tool	Seconds	Someone else	Medium — the exposed settings	One-off clips, testing prompts, no install
Serverless API	An API key	Provider (fal, WaveSpeed and others per the model repo)	High, via parameters	Automation, batches, app integration
Hugging Face Space	Seconds	Community GPU queue	Low to medium	Free exploration, comparing checkpoints
Local install	An afternoon	You	Total	Volume work, LoRAs, privacy, no per-clip cost

The honest summary: browser tools are for deciding whether Wan 2.2 suits your material. APIs are for doing it a hundred times. Local is for doing it a thousand times, or for anything you cannot upload.

Route 1: A Hosted Browser Tool

This is what most people mean by "Wan 2.2 online". You open a page, type a prompt or upload a still, press generate, and get an MP4. No Python, no checkpoints, no VRAM anxiety.

What it is good for is speed of iteration. The thing that makes Wan 2.2 hard to evaluate from benchmarks is that quality depends heavily on prompt and settings, and you can only learn that by generating twenty variants of the same idea. A hosted tool makes each variant cost seconds instead of a download.

What it is not good for is fine control. Distilled step counts, LoRA stacks, samplers, block swapping — the settings that separate a decent Wan 2.2 clip from a great one are exactly the ones a simplified UI hides. That is not a flaw in the tool; it is what makes it usable by people who do not want to learn ComfyUI.

If you want to see the model's character before committing to any pipeline, this is the cheapest way in: the free Wan 2.2 video generator runs text-to-video and still-to-motion without an account ritual, and the Wan 2.2 model overview explains which variant sits behind the button.

If you want output today, start here: Launch Wan 2.7 Now →

Route 2: Serverless Inference APIs

The official model repositories list third-party inference providers — fal and WaveSpeed appear on the Wan2.2 model cards — which means you can call Wan 2.2 without owning a GPU and without a web form.

This is the route for anything programmatic:

• Batch generation where you need 200 variants overnight
• Product integration, where a user action triggers a render
• Benchmarking your own prompts against a fixed checkpoint so results are comparable over time

The trade-off is that you inherit the provider's model version and defaults, not the upstream repo's. When a provider pins TI2V-5B at 24fps and 720p, you get that and not the pip-installable knobs. Pricing is set per provider and per second of output — check the model page before you build a cost model around it, because per-second video pricing varies more between providers than text pricing does between LLM hosts.

One practical caveat: serverless endpoints are usually running the base checkpoint. Community LoRAs, which are a large part of why people still choose Wan 2.2, are generally not available there. If your look depends on a specific LoRA, this route is closed to you unless the provider supports custom weights.

Route 3: Hugging Face Spaces

The Wan2.2-TI2V-5B repository lists 100 Spaces built on that checkpoint alone, plus 137 finetunes and 16 quantised builds in its model tree. That ecosystem is the strongest argument for Wan 2.2 in 2026: whatever style you want, someone has probably trained an adapter for it.

Spaces are free to try and often queue-bound. Two things to know before you rely on one:

They are community-maintained. Uptime, GPU class and whether the Space is running the base model or a LoRA merge are all up to the author. A Space that produced excellent output last month may be a different build today.

You get less control than an API. Most Spaces expose prompt, image, size and maybe a seed. That is enough to judge a checkpoint, not enough to run production.

Used properly, Spaces are a comparison tool. Run the same prompt through a T2V-A14B Space and a TI2V-5B Space and you will see the actual difference between the 14B MoE pair and the 5B hybrid — which is a difference in motion complexity more than in sharpness.

Animate a Still with Wan 2.2

Route 4: Local, and When It Beats Hosting

Local is the fourth option and, for volume work, the cheapest. The published requirements are specific enough to plan against:

Checkpoint	Type	Resolution	Single-GPU inference
T2V-A14B	MoE, text-to-video	480P & 720P	~80GB VRAM for the documented single-GPU command; offload flags reduce the peak
I2V-A14B	MoE, image-to-video	480P & 720P	~80GB class; offload flags required on consumer cards
TI2V-5B	Hybrid T2V+I2V, 16×16×4 VAE	720P @ 24fps	24GB is enough — the documented command runs on a 4090
S2V-14B	Speech-to-video	—	Separate checkpoint; consumes audio, does not generate it

The 5B is the important row. A 24GB consumer card generating 720p at 24fps is why Wan 2.2 stayed relevant after newer models arrived: the hardware floor is a gaming PC, not a server.

Local also unlocks the parts of the ecosystem that online routes cannot give you — LoRA stacks, quantised GGUF builds, ComfyUI graphs, and the ability to keep client footage off someone else's disk. If you want the local path in detail, the Wan 2.2 I2V 14B Fast guide covers the distilled 4-step preset and its VRAM tiers, and the Wan 2.2 free options breakdown compares the no-cost hosted routes.

Which Wan 2.2 Checkpoint Do You Actually Need?

Four checkpoints, four jobs. Getting this wrong wastes more time than any prompt mistake:

If your input is…	Use	Why
Text only	T2V-A14B, or TI2V-5B for speed	The A14B pair renders more complex motion; the 5B is faster at 720p
A still image	I2V-A14B, or TI2V-5B with an image	The 5B's ti2v task switches to image conditioning when you pass an image
An image plus a voice track	S2V-14B	The only 2.2 checkpoint that consumes audio
A style you trained	Whichever base your LoRA targets	Most community LoRAs are built on 2.2 bases

Two details from the official repo that save real debugging time:

In ti2v-5B, passing an image silently switches the task. The documentation is explicit: if the image parameter is configured, it is image-to-video; otherwise it defaults to text-to-video. There is no separate toggle to forget.

Ready to try it yourself? Try Wan 2.7 Free →

Aspect ratio is inherited. In the 5B image-to-video path, the size parameter sets the area of the output and the aspect ratio follows the input image. So a portrait still gives you a portrait clip — you do not fight a 16:9 default.

Settings That Decide Whether the Output Is Usable

Three settings do most of the work on Wan 2.2, and they are the three most often left at defaults:

Resolution and size. For the 720p TI2V task the documented resolution is 1280×704 (or 704×1280), not 1280×720. Matching the model's trained size avoids a rescale that softens the frame before you have even looked at it.

Prompt extension. The official repo recommends enabling prompt extension, because a short prompt leaves detail on the table. It ships two implementations — a DashScope call (documented with qwen-plus for text-to-video and qwen-vl-max for image-to-video) or a local Qwen model. The practical consequence: on a hosted tool you get whatever the platform enabled; on an API or local run, this is a setting you can actually choose. Note that extension rewrites your prompt, so for precise prompt engineering, test with it off to see the raw response.

Steps and duration. Wan 2.2's standard output is around five seconds with a 24fps 720p target on the 5B. Distilled schedules (Lightning LoRAs) cut steps hard and change character — crisper drafts, weaker large-motion resolution. If your clips look "almost right but weird in the second half", the step schedule is a likelier culprit than your prompt. For a closer look at how it stacks up against other models, see LTX 2.3 vs Wan 2.7. If you want to test it without installing anything, the free Wan video generator works in the browser.

Cost Reality, Route by Route

Route	What you pay	Hidden cost
Hosted browser tool	Free tier, or a plan if you need more volume	Queue position; simplified settings
Serverless API	Per second of generated video, set by the provider	Model version pinned by the provider; LoRA support rare
Hugging Face Space	Free	Queue time; zero support; build changes without notice
Local (TI2V-5B)	Hardware + electricity	An afternoon of setup; ~24GB VRAM
Local (A14B pair)	Hardware + electricity	80GB-class VRAM or aggressive offload; slower per clip

The economics flip at predictable points. If you generate fewer than a dozen clips a month, hosted is free and local is a waste of an evening. If you generate hundreds, per-second API pricing adds up fast and a 4090 pays for itself. In between, a free browser tool plus occasional API calls covers it.

Generate Wan 2.2 Clips Without a GPU

When Wan 2.2 Is the Right Call in 2026 — and When It Isn't

Stay on Wan 2.2 if:

• You need free and fast more than you need the last 20% of quality
• Your look depends on a community LoRA trained on 2.2 checkpoints
• You are prototyping motion ideas where 720p is plenty
• You want a browser session, not a pipeline

Move to Wan 2.7 if:

• You need audio — 2.2's base checkpoints are silent, full stop
• You need 1080P output rather than 720P
• You need multi-shot narrative inside one clip, or clips up to 15 seconds instead of five
• You need native prompt following on complex, multi-subject scenes

Want to see the difference on your own footage? Start creating with Wan 2.7 →

The generational gap is real but narrow in one direction: 2.7 is dramatically better at sound, duration and instruction-following, and only moderately better at a single clean five-second 720p shot of a static subject. If that single shot is your entire use case, the upgrade buys you very little.

A full comparison of Wan 2.2 and Wan 3.0 covers the newer generation side by side, and what Wan 2.2 is and how it fits the family is the place to start if you are new to the model line.

The Bottom Line

Wan 2.2 online comes down to one question: are you evaluating the model or producing with it? Browsers and Spaces are for evaluating — free, instant, limited. APIs are for producing at moderate scale with automation. Local is for producing at high volume, or when the footage cannot leave your machine.

Pick the checkpoint by input type — text to T2V, stills to I2V or the 5B's image path, voice to S2V — and remember that 720p, five seconds and silence are the model's actual specification, not settings you can talk it out of.

Try Wan 2.2 for Free

No GPU, no checkpoint download, no queue on a stranger's Space.

Wan 2.2's 5B checkpoint is the reason a browser tool can produce 720p video at 24fps at all — so use that instead of installing 30GB of weights to find out whether the model suits your material:

• Text-to-video and image-to-video in one interface, no separate models to pick
• 720p output matching the model's trained resolution
• Free to try, so prompt experiments cost time rather than credits
• No local GPU required — the same 24GB-class work happens on our hardware

Generate a Wan 2.2 Clip Free

And when your project outgrows silent five-second clips, Wan 2.7 on Pollo is the same model family with audio, 1080P and longer durations — the upgrade path for work that has to ship rather than test.

FAQ

Is Wan 2.2 still available online in 2026?

Yes. The official weights are open under Apache 2.0 and remain on Hugging Face, and both hosted tools and third-party inference providers still serve the checkpoints. It is a previous generation, not a retired one.

Can I run Wan 2.2 in a browser without a GPU?

Yes. Hosted tools run the model on their hardware and return an MP4, so nothing is installed locally. The TI2V-5B checkpoint is the practical choice for this because it was designed to run on a single 24GB consumer card.

Does Wan 2.2 generate audio?

No. T2V-A14B, I2V-A14B and TI2V-5B all output silent video. Audio handling in Wan 2.2 exists only in S2V-14B, which takes an existing audio track and drives a performance from it — it does not create sound.

What resolution does Wan 2.2 online actually produce?

720p is the ceiling. The TI2V-5B text-image-to-video task is documented at 1280×704 (or 704×1280) at 24fps, and the A14B pair supports 480P and 720P. Anything above 720p requires an upscale afterwards.

How long are Wan 2.2 clips?

The A14B checkpoints are documented around five-second output, and the 5B targets five seconds at 720p/24fps. For longer or multi-shot work, Wan 2.7 supports clips up to 15 seconds with native multi-shot output.

Which Wan 2.2 checkpoint should I use for image-to-video?

I2V-A14B is the dedicated image model. The TI2V-5B is often the better practical pick because it handles both text and image inputs in one checkpoint and runs on far less VRAM — pass it an image and it switches to image conditioning automatically.

Is Wan 2.2 good enough for commercial work?

The weights are Apache 2.0 and the Wan team states no claim over generated content, so commercial use is permitted under the licence. Quality-wise, 720p silent five-second clips suit social cuts, product loops and b-roll; they are thin for anything that needs sound or a hero frame above 720p.

Related guides

• What Wan 2.2 is, and where it sits in the Wan family
• Wan 2.2 I2V 14B Fast: what the 4-step preset changes
• Wan 2.2 free and unlimited: every no-cost route compared
• Running Wan 2.2 through Perchance
• Wan 2.2 vs Wan 3.0: what actually changed

References

• Wan2.2 model card — TI2V-5B (Hugging Face, Wan-AI)
• Wan2.2 model card — T2V-A14B (Hugging Face, Wan-AI)
• Wan2.2 model card — I2V-A14B (Hugging Face, Wan-AI)
• Wan2.2 inference code repository (GitHub, Wan-Video)
• Wan: Open and Advanced Large-Scale Video Generative Models (arXiv:2503.20314)
• Wan official site