Wan2.2
Open video generation models for text-to-video, image-to-video, speech-to-video and character animation.
Wan2.2 offers text, image, speech and animation video models with a 5B model for a 24GB card.
Wan2.2 is a family of open video generation models with inference code, from Wan-Video. It suits teams with GPUs who want to generate video locally or build on open weights rather than rent a hosted generator.
What it is
The repo ships inference code and checkpoints for text-to-video and image-to-video (A14B Mixture-of-Experts models at 480P and 720P), a 5B text-image-to-video model at 720P, an audio-driven Speech-to-Video model and Wan2.2-Animate for character animation and replacement. It is integrated with ComfyUI and Diffusers.
Who it's for
It suits ML engineers, studios and researchers with serious GPU hardware. It does not suit editors who want a timeline app, and the 14B models are out of reach without an 80GB card or several GPUs.
Key features
Mixture-of-Experts video model
A14B models split denoising across timesteps between specialised experts to enlarge capacity.
Text and image to video
T2V-A14B and I2V-A14B generate 480P and 720P clips.
Compact 5B hybrid model
TI2V-5B handles text-to-video and image-to-video at 720P with a high-compression VAE.
Speech-to-video
S2V-14B generates audio-driven video, with optional CosyVoice text-to-speech.
Character animation
Animate-14B animates and replaces characters while copying movement and expression.
ComfyUI and Diffusers
Integrations make the models usable inside common open-source pipelines.
Up and running in 3 steps
Our honest verdict
What's great
- Open weights and code mean you can generate video on your own hardware with no per-clip fee.
- The 5B model runs on a single GPU with at least 24GB VRAM, such as an RTX 4090, per the README.
- Several tasks in one family: text, image, speech and character animation.
- Community projects add speed-ups, including low-VRAM offload, FP8 quantization and LoRA training through DiffSynth-Studio.
Where it falls short
- The 14B text and image models need a GPU with at least 80GB VRAM for single-GPU inference, or multi-GPU setups with FSDP and DeepSpeed.
- Setup is a Python and PyTorch install (torch 2.4 or later, flash_attn can fail to build) plus large checkpoint downloads.
- It is a command-line generation script, not an editor, so there is no timeline, trimming or audio mixing.
- The licence text says you are accountable for your use and must not use outputs to break laws, cause harm, spread misinformation or target vulnerable people.
Pricing & license
The code and models are free under Apache-2.0, and the only cost is the GPU hardware or rented compute you run them on. The README links a hosted site at wan.video, but no price could be read there.
Open remote AI-training roles
870 open remote AI-training roles, updated November 14, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
What GPU do I need?
At least 24GB VRAM for TI2V-5B and at least 80GB for the A14B models on a single GPU, per the README.
Can I use the output commercially?
The README says the models are Apache 2.0 and the project claims no rights over your generated content, but you must follow the licence restrictions on harmful or unlawful use.
Is there a hosted version?
The README links a hosted site at wan.video, but its pricing could not be read from the page, so check the site directly.
