Generate AI video with Wan 2.2: choose the right model variant, run it locally or via a hosted API, prompt for cinematic control, and ship finished clips.
Wan 2.2 is Alibaba's open-source video generation family, released in late July 2025 under an Apache-2.0 license. The headline is that it is open: you can download the weights from Hugging Face, GitHub, or ModelScope and run them yourself, and the smallest variant was built to render 720p on a single consumer graphics card. For a creator that means a capable text-to-video and image-to-video model with no per-clip meter — if you can run it, or if you reach it through a hosted API.
This is the practical walkthrough: how to actually get a clip out of Wan 2.2. It covers picking the right variant for your job, deciding whether to run it locally or through a provider, writing prompts that use the model's cinematic controls, and — the step most tutorials skip — turning the raw 5-second clip into something finished and published. Where a spec depends on which variant you run or which host you use, this says so instead of pretending there is one number; verify the current details on the model's own repo before you build a batch around them.
The steps
Pick the Wan 2.2 variant that fits your job and your hardware. Wan 2.2 ships three models. T2V-A14B generates video from a text prompt and I2V-A14B animates a still image; both are Mixture-of-Experts models with about 27B total parameters (roughly 14B active per step), and Alibaba recommends serious VRAM — on the order of 80GB — for single-GPU inference. TI2V-5B is the dense hybrid that does both text- and image-to-video and was designed to run 720p on a single 24GB consumer card like an RTX 4090. If you have a normal workstation, the 5B is your model; the A14B pair is for a rented high-VRAM GPU or a hosted endpoint.
Decide how to run it: local weights or a hosted API. Open weights do not mean zero cost — they mean the cost is your hardware and setup instead of a per-clip bill. Running locally gives you unlimited generations and full privacy but needs a capable GPU and some configuration. A hosted provider (Alibaba Cloud Model Studio, or third-party endpoints that serve Wan) removes the setup and bills per generation or per second of output, which is cheaper for a handful of clips and more expensive at volume. Pick local if you will generate a lot or need the footage private; pick a host to test the model before committing hardware. Confirm current rates on the provider's page — they vary by resolution and change often.
Install the model (ComfyUI or Diffusers) if you are running it yourself. Wan 2.2 integrates with ComfyUI and Hugging Face Diffusers, which are the two realistic local paths. Accept the model terms and download the variant you chose from Hugging Face or ModelScope, then load the matching workflow or pipeline. ComfyUI gives you a node graph you can save and reuse per project; Diffusers suits anyone scripting generations in Python. Do a single short test render before any batch so you catch a VRAM or dependency problem on one clip instead of fifty.
Write a prompt that uses Wan 2.2's cinematic controls. Wan 2.2 was trained on curated cinematic-aesthetic data, so prompt language about lighting, time of day, color tone, camera angle, composition, frame size, and focal length actually steers the output rather than being ignored. Write the subject and action first, then add the shot: "golden-hour backlight, low angle, shallow focus, cool color grade." Keep one clear action per clip — the model generates about 5 seconds per run, so a prompt that asks for three beats will rush all of them. Treat each generation as one shot, not a scene.
For image-to-video, prep a strong source frame first. With I2V-A14B or the 5B, the still you feed in decides almost everything — the clip inherits its composition, lighting, and subject. Use a clean, high-resolution image with the subject clearly separated from the background, and describe the motion you want in the prompt ("slow push-in, hair moving in the wind") rather than re-describing the scene the image already shows. A weak or busy source image produces a weak clip no prompt can rescue, so spend the time on the frame.
Generate, then iterate on seed, length, and resolution. Run the generation and judge the first result against the shot you wanted, not the first thing that renders. Change the seed to get a different take on the same prompt, nudge the prompt wording for motion or framing problems, and remember the per-clip ceiling is short — around 5 seconds at 720p, and Alibaba only documents a frame rate for the 5B (24fps) — confirm the A14B output's frame rate on the model card before mixing it with other clips. Lock a prompt and seed that work before you scale, so a batch is reproducible instead of a lottery.
Export, then stitch or extend if you need length. Export the clip at the resolution the variant produced (720p is the realistic target across the line). Because each generation is only a few seconds, a longer piece is several clips assembled on a timeline — or an image-to-video chain where the last frame of one clip seeds the next. Decide the frame rate you are delivering at and keep it consistent across clips so a stitched sequence does not stutter where two renders meet.
Caption, reframe, and publish the clip. A raw Wan 2.2 clip is not a post yet. Feeds autoplay muted, so add captions; each platform wants a different aspect ratio, so reframe 16:9 to 9:16 or 1:1 where needed; and a single 5-second clip usually needs to be cut, combined, or wrapped with a hook and a call to action before it earns attention. This finishing-and-distribution step is separate from the model entirely — Wan 2.2 hands you footage, and turning footage into published content is its own workflow.
Common gotchas
The A14B models are not consumer-GPU models. They want roughly 80GB of VRAM for single-GPU inference — if you are on a normal card, the TI2V-5B is the variant that fits (720p on a 24GB GPU).
Generations are short. Each run produces about a 5-second clip, so plan for stitching or image-to-video chaining if you need real length, and do not prompt for a whole scene in one go.
Frame rates differ by variant. Alibaba documents 24fps for the TI2V-5B; the A14B models' output frame rate isn't stated on the official model card — confirm it on your own export before mixing clips on a timeline.
"Open weights" is not "free to run." You supply the GPU, the setup, and the electricity; for low volume a hosted API can be cheaper in total effort even though it charges per clip.
Hosted pricing varies a lot. Per-second and per-generation rates differ by provider and resolution and change often — confirm the current number before a batch run, not after.
The model stops at the clip. There is no captioning, reframing, voiceover, or publishing in Wan 2.2 — everything after the MP4 is another tool's job.
Legal note
Wan 2.2's Apache-2.0 license permits commercial use of the model, but that does not clear what you generate: a prompt that recreates a real person's likeness, a trademarked character, or copyrighted footage can still create rights problems regardless of the license. Separately, most platforms now require you to disclose AI-generated or synthetic video — label it under each platform's AI-content rules when you publish.
Where Kompozy fits
Follow the steps and you will notice the shape of the work: steps 1 through 7 get you a 5-second clip, and step 8 — caption it, reframe it for each feed, give it a hook, publish it — is a whole separate job that repeats for every clip you ever make. Wan 2.2 is genuinely good at the generation half and does nothing for the finishing half. That is the gap [Kompozy](/) is built to close, so a folder of raw Wan renders becomes posts instead of footage.
Drop a Wan 2.2 clip into Kompozy as a source and it takes over at step 8: [Clipped Shorts](/glossary/clipped-short) cut and reframe the footage to 9:16, 1:1, or 16:9 with word-synced captions burned in, so one clip ships to every feed in the right shape without you touching an editor. Kompozy also fans the same idea wider into a [Carousel](/glossary/hyperframes), quote graphics, a blog, and a newsletter from the same source. On the weeks you do not render anything in Wan, it generates net-new video instead of leaving a gap: avatar-led [Persona Shorts](/glossary/persona-shorts) keep the channel alive. A single [Persona Brief](/glossary/persona-brief) holds one voice across every output so your Wan clips and everything around them read as the same creator, and [Autopilot](/glossary/autopilot) schedules and publishes the set across the eight social platforms plus blog and email behind a per-post review gate. The honest boundary: Kompozy does not generate the cinematic footage — that is Wan 2.2's job — and it will not run your local GPU setup. It takes the clip the moment it exists and does the eighth step, at scale. Starter ($199/mo, 5,500 credits) fits a solo creator turning a few Wan clips a week into a full posting schedule; Pro ($499/mo, 18,000 credits) suits teams running high-volume, multi-format publishing; Enterprise is custom via Contact Sales.
Frequently asked questions
Can I run Wan 2.2 on my own computer?
The TI2V-5B variant was built to generate 720p video on a single 24GB consumer GPU such as an RTX 4090, so yes on a capable workstation. The larger A14B text-to-video and image-to-video models need far more VRAM (Alibaba suggests around 80GB for single-GPU inference), so those are better run on a rented high-memory GPU or a hosted endpoint.
Is Wan 2.2 free to use?
The weights are open under an Apache-2.0 license and free to download from Hugging Face, GitHub, or ModelScope, so there is no license fee. Running it still costs your hardware, setup, and electricity, and hosted providers that serve Wan 2.2 bill per generation or per second of output — confirm current rates on the provider's page.
How long can a Wan 2.2 clip be?
Each generation produces a short clip — around 5 seconds at 720p across the line. For anything longer you assemble several clips on a timeline, or chain image-to-video so the final frame of one clip seeds the next. There is no single-pass long-form render in Wan 2.2.
What is the difference between the A14B and 5B models?
The A14B models (T2V and I2V) are Mixture-of-Experts with about 27B total parameters and the higher quality ceiling, but they need heavy VRAM. The TI2V-5B is a smaller dense model that handles both text- and image-to-video and runs 720p on a single consumer card at 24fps — the practical choice if you are generating locally on normal hardware.
Does Wan 2.2 add captions or publish the video for me?
No. Wan 2.2 generates footage and nothing else — there is no captioning, reframing per platform, voiceover, or scheduling. Turning the raw clip into a finished, captioned, published post is a separate workflow you run after the model hands you the file.