How to Use Wan 2.2: Text and Image to Video Online or Locally (2026)

To use Wan 2.2, pick text-to-video or image-to-video, write a short prompt that names the subject, the scene and the motion, choose a length and resolution, and generate. You can do this in a browser studio such as Kenerate AI, where Wan 2.2 makes 5 or 8 second clips at 480p or 720p with no GPU, or run the open weights yourself from Alibaba's Wan2.2 GitHub repo. Locally, the small TI2V-5B model needs a 24 GB GPU and the 14B models need about 80 GB on a single card.

Wan 2.2 is Alibaba's open-weights video model family, licensed under Apache 2.0. This guide covers both routes step by step, the prompt structure Alibaba recommends, and the mistakes that cost the most re-renders. Model facts come from the Wan2.2 GitHub README and Alibaba Cloud's prompt guide, read on 23 October 2026. Kenerate facts come from our video model config. Kenerate AI is our product.

What is Wan 2.2, and which version do you need?

Wan 2.2 is not one model but five. The two 14B "A14B" models use a mixture-of-experts design: a high-noise expert lays out each clip in the early steps and a low-noise expert refines the detail later. The README puts that at about 27B parameters in total with 14B active per step.

For most people the choice is simple. If you have a 24 GB card, start with TI2V-5B. If you don't have a GPU, use a hosted version.

How do the ways to use Wan 2.2 compare?

How do you use Wan 2.2 on Kenerate AI, step by step?

Kenerate runs two versions: Wan 2.2 and Wan 2.2 Ultra Fast. Video is a paid-plan feature on Kenerate, so you need a credit pack first. Packs start at $15 for 1,300 credits on the pricing page.

1. Open the video studio.
2. Choose Text to Video or Image to Video.
3. Pick Wan 2.2 or Wan 2.2 Ultra Fast in the model list.
4. For image to video, upload your start frame. A prompt is required in this mode, so describe the motion you want.
5. Write the prompt (see the formula below).
6. Choose 5 or 8 seconds, 480p or 720p, and 16:9 or 9:16.
7. Generate. Clips come back silent, so add music or a voice-over in your editor if you need sound.

Credits come from our video model config, per second of video:

A sensible workflow: draft the prompt with Ultra Fast at 480p for 15 credits, and only render the final with Wan 2.2 at 720p once the motion is right. Ultra Fast is the cheapest video model Kenerate offers. Our Wan 2.2 page shows real clips from both versions.

How do you run Wan 2.2 locally from the GitHub repo?

These steps follow the official Wan2.2 README.

1. Clone the repo: git clone https://github.com/Wan-Video/Wan2.2.git, then cd Wan2.2.
2. Install dependencies: make sure torch >= 2.4.0, then pip install -r requirements.txt. The README notes that if flash_attn fails, you should install the other packages first and flash_attn last.
3. Download the weights: pip install "huggingface_hub[cli]", then for example huggingface-cli download Wan-AI/Wan2.2-TI2V-5B --local-dir ./Wan2.2-TI2V-5B.
4. Generate with the 5B model on one GPU:

bashCopy
python generate.py --task ti2v-5B --size 1280*704 --ckpt_dir ./Wan2.2-TI2V-5B \
  --offload_model True --convert_model_dtype --t5_cpu \
  --prompt "A red fox trots through fresh snow in a birch forest, snowflakes falling, slow tracking shot"

5. For image to video with the 5B model, add --image path/to/start.jpg to the same command.
6. For the 14B models, use --task t2v-A14B or --task i2v-A14B with --size 1280*720, plus --offload_model True --convert_model_dtype. The README says these single-GPU commands need at least 80 GB of VRAM.

The README says TI2V-5B can make a 5-second 720P video in under 9 minutes on a single consumer GPU. Note the size: for the 5B model at 720P, use 1280*704 or 704*1280, not 1280*720.

The flags --offload_model True, --convert_model_dtype and --t5_cpu all reduce GPU memory. If you run out of memory, add them. If you have an 80 GB card, the README says you can drop them for speed.

How do you run Wan 2.2 in ComfyUI?

ComfyUI ships Wan 2.2 templates, which is the easiest local route if you don't want to use the command line.

1. Update ComfyUI to the latest version.
2. Open Workflow Templates, go to the Video category and pick a Wan 2.2 template (text to video, image to video or the 5B model).
3. Download the model files the template lists into the matching models folders.
4. For image to video, load your start frame in the image node.
5. Edit the prompt, set the size and length, and click Run.

Quantized (fp8) files in the templates lower the VRAM you need, at some cost to quality.

How should you structure a Wan 2.2 prompt?

Alibaba Cloud's prompt guide for Wan gives three formulas:

• Basic: Entity + Scene + Motion.
• Advanced: Entity (description) + Scene (description) + Motion (description) + Aesthetic control + Stylization.
• Image to video: Motion + Camera movement. The image already shows the subject and scene, so spend the words on what moves and how the camera moves.

Put together: "A red fox with a thick winter coat trots through a quiet birch forest in fresh snow, snowflakes falling, then stops and looks to the left. Soft winter daylight, medium shot, slow tracking shot, wildlife documentary style."

The guide's camera vocabulary includes "camera pushes in", "camera pulls out", "camera tilts up", "fixed camera", orbiting and tracking shots. Its lighting words include soft light, backlight, rim light and side light. For 18 copy-ready examples, see our Wan 2.2 image-to-video prompt guide.

What are the most common Wan 2.2 mistakes?

1. Re-describing the image in image to video. The start frame already shows the subject. Describe motion and camera movement instead.
2. Too many actions. In 5 to 8 seconds, one main action plus one camera move reads better than four events.
3. Using 1280*720 with the 5B model. The README asks for 1280704 or 7041280 at 720P.
4. Running the 14B models on a small card. Without about 80 GB of VRAM, use the 5B model, the offload flags, quantized files or a hosted version.
5. Expecting sound. Wan 2.2 text and image to video makes silent clips. Speech-driven video is a separate model (S2V-14B).
6. Final-quality renders while testing. Test at 480p or with a faster version, then render the keeper at 720p.
7. Conflicting directions. "Fixed camera" plus "tracking shot" gives the model two jobs. Pick one.

When should you use a newer Wan model instead?

Wan 2.2 is the cheapest Wan to run and the best documented for local use. If you need sound, longer clips or 1080p, a newer version fits better. Kenerate also runs Wan 2.5, 2.6, 2.7 and 3.0. We compare them in Wan 2.2 vs Wan 2.7 vs Wan 3, and our Wan model page lists every version. To transfer motion from a video onto a character, see Wan 2.2 Animate.

Bottom line

Using Wan 2.2 comes down to three choices: where to run it, which mode, and how to prompt. Without a GPU, a browser studio is the fastest start. On Kenerate a 5-second Ultra Fast draft costs 15 credits on a paid plan. With a 24 GB GPU, the official TI2V-5B command or a ComfyUI template gets you running locally. In both cases, write Entity + Scene + Motion for text to video and Motion + Camera movement for image to video, keep to one main action, and test cheap before you render the final.