Wan 2.2 image-to-video.
Wan 2.2 image-to-video is an open-source video model from Alibaba (Wan-AI), coming soon to Windpaint. Its planned price is 4 credits ($0.20) per 5 second clip at 480p. Request early access to start with the live models.
- Status
- Coming soon
- Task
- Image to video
- Input
- Prompt and start frame
- Output
- Video clips
- Developer
- Alibaba (Wan-AI)
- Parameters
- 27B MoE, 14B active
- Licence
- Apache-2.0
- Released
- July 2025
What the model does.
Paraphrased from the official model card.
Wan 2.2 I2V A14B is Alibaba's open-weight image-to-video model. It animates a start image you supply, guided by a text prompt describing the motion.
It uses a mixture-of-experts design: two experts of about 14B parameters each, one for the early high-noise steps and one for the later detail steps, so 27B in total with 14B active per step. The official code supports 480p and 720p output.
Coming soon to Windpaint.
Wan 2.2 image-to-video isn't live yet. Its planned price is below; when it goes live, this page gets its request fields like the live models have.
- 480p, 5 s (coming soon)
- 4 credits ($0.20) per clip
Want this model? Request early access. We're adding models as we add capacity.
Is Wan 2.2 image-to-video licensed for commercial use?
Licence: Apache-2.0.
Yes. Apache-2.0 allows commercial use, including running the model behind a paid API, with no revenue or user caps.
- Apache 2.0. The Wan team states it claims no rights over generated content. Wan2.2 repo
- The README (not the Apache licence text) adds that use must not involve sharing content that violates applicable laws, harms individuals or groups, disseminates personal information intended for harm, spreads misinformation, or targets vulnerable populations. Wan2.2 repo
This is our reading of the licence, not legal advice. Read the licence text before you rely on it.
Wan 2.2 image-to-video VRAM requirements.
Published by Alibaba (Wan-AI) for running the weights yourself. These are the developer's figures, not Windpaint's.
- Published VRAM
- At least 80 GB
- Applies to
- Single-GPU 720p command with model offload and dtype conversion. Multi-GPU example uses 8 GPUs.
- The README's single-GPU I2V-A14B command (
--size 1280*720 --offload_model True --convert_model_dtype) can run on a GPU with at least 80GB VRAM. Wan2.2 repo - For out-of-memory errors the README suggests
--offload_model True,--convert_model_dtypeand--t5_cputo reduce GPU memory use. Wan2.2 repo - Multi-GPU inference uses PyTorch FSDP plus DeepSpeed Ulysses; the official example runs on 8 GPUs (
torchrun --nproc_per_node=8 --dit_fsdp --t5_fsdp --ulysses_size 8). The README's GPU tests used FlashAttention3 on Hopper GPUs. Wan2.2 repo - Two experts, a high-noise expert for early steps (overall layout) and a low-noise expert for later steps (detail), each about 14B parameters: 27B in total, 14B active per step. Wan2.2 repo
- The Wan team publishes Diffusers-format weights as
Wan-AI/, and the card lists ComfyUI integration as done. Wan2.2-I2V-A14B-Diffusers model cardWan2.2-I2V-A14B-Diffusers
Wan 2.2 image-to-video native resolution.
What the developer's own code and examples use.
- Accepted
--sizevalues for i2v-A14B:1280*720,720*1280,832*480,480*832(720P and 480P). Wan2.2 repo, __init__.py - For image-to-video in the official code,
--sizesets the area of the output video and the aspect ratio follows the input image. Wan2.2 repo - Shared config defaults:
frame_num=81,sample_fps=16. Wan2.2 repo, shared_config.py
Recommended settings for Wan 2.2 image-to-video.
Steps, guidance and defaults from the developer's card and code, for running the weights yourself.
- Official I2V-A14B config: 40 sampling steps, sample_shift 5.0, expert switch boundary 0.900, guidance scale 3.5 for both the low-noise and high-noise experts. Wan2.2 repo, wan_i2v_A14B.py
- Diffusers example:
max_area480*832 with height and width derived from the input image's aspect ratio,num_frames=81,guidance_scale=3.5,num_inference_steps=40, exported atfps=16. Wan2.2-I2V-A14B-Diffusers model card - The shared config ships a default Chinese negative prompt. Its terms include oversaturated colours, overexposure, static, blurry detail, subtitles, overall grey, worst and low quality, JPEG artifacts, extra or fused fingers, badly drawn hands and faces, deformed limbs, motionless frames, messy background, three legs, many people in the background, and walking backwards. Wan2.2 repo, shared_config.py
Wan 2.2 image-to-video prompt guide.
From Alibaba (Wan-AI)'s own prompt guide, prompt tooling and examples.
- The Wan team recommends prompt extension (rewriting the prompt with an LLM) to enrich detail. For image-to-video they suggest qwen-vl-max via Dashscope, or a local Qwen2.5-VL-7B-Instruct or Qwen2.5-VL-3B-Instruct. Wan2.2 repo
- The official I2V rewrite instructions: describe motion rather than static scene content already visible in the image, keep the subject's actions, keep and emphasise camera moves such as "the camera pans up" or "the camera moves from left to right", and stay at 100 words or less. Wan2.2 repo, system_prompt.py
- If the user gives only an action (for example "dancing"), the official rewrite instructions fill in the subject from the image ("a girl is dancing"). Wan2.2 repo, system_prompt.py
Wan 2.2 image-to-video limitations.
As the developer states them.
generate.pysays the frame count should be 4n+1. Wan2.2 repo, generate.py- Only four sizes are accepted for i2v-A14B:
1280*720,720*1280,832*480,480*832. Wan2.2 repo, __init__.py
Questions.
Can I call Wan 2.2 image-to-video on Windpaint today?
Not yet. Wan 2.2 image-to-video is coming soon. Its planned price is 4 credits ($0.20) per 5 second clip at 480p. There's no date for Wan 2.2 image-to-video.
How much VRAM does Wan 2.2 I2V need?
The official README says its single-GPU 720p command for I2V-A14B, with model offloading and dtype conversion on, can run on a GPU with at least 80GB of VRAM. It also documents multi-GPU inference, with an 8-GPU example.
What resolution does Wan 2.2 I2V support?
480p and 720p. For image-to-video the size setting (for example 1280*720) sets the pixel area of the output, and the aspect ratio follows the input image.
What does A14B mean in Wan 2.2 I2V A14B?
It's a mixture-of-experts model with two experts of about 14B parameters each: a high-noise expert for the early denoising steps and a low-noise expert for the later detail steps. That's 27B parameters in total, with 14B active per step.
Is there a Wan 2.2 5B image-to-video model?
Yes. Wan2.2-TI2V-5B does both text-to-video and image-to-video at 720p and 24 fps.
What negative prompt does Wan 2.2 I2V use?
The official config ships a default negative prompt in Chinese. It includes terms for oversaturation, overexposure, static or motionless frames, blur, subtitles, low quality, JPEG artifacts, malformed hands, faces and limbs, messy or crowded backgrounds, and walking backwards.
Can Wan 2.2 be used commercially?
The models are licensed under Apache 2.0, and the Wan team says it claims no rights over the content you generate. The README asks that use not involve sharing unlawful or harmful content.
Is Wan 2.2 image-to-video open source?
Yes. Its weights are licensed Apache-2.0, an OSI-approved open-source licence.
How do I ask for Wan 2.2 image-to-video?
Request early access at app.windpaint.ai/register. We're letting people in gradually while we add models and capacity.