What Wan 2.2 image-to-video does
Wan 2.2 is the 2025 generation of Alibaba's open video models, released in July 2025 under the Apache 2.0 licence. The image-to-video variant, Wan2.2-I2V-A14B, uses a mixture-of-experts diffusion transformer: one expert handles the early, noisy steps that decide the overall motion, another handles the late steps that refine detail. About 14 billion parameters are active per step out of 27 billion in total. A smaller hybrid model, Wan2.2-TI2V-5B, accepts both text and an image and runs on a single consumer card.
You give the model a picture, the first frame of the clip, and a prompt that says what should move. It generates 81 frames at 16 frames per second, about five seconds, and keeps the identity of the subject far better than earlier open models: faces, clothes and lighting stay consistent while the camera or the subject moves.
How to write the motion prompt
The image already carries the subject, the style and the composition, so the prompt only needs to describe the change. Write one or two short sentences: who or what moves, how, and what the camera does. "The woman turns her head to the right and smiles, soft wind in her hair, slow push-in" works; a list of style tags does not. Mention the camera explicitly when it matters: "static camera", "slow pan to the left", "the drone rises".
Keep the motion plausible for a five-second clip. One action per prompt animates cleanly; three actions in a row produce jerky or truncated results. Verbs such as walks, waves, blinks, flows, drifts, rotates give predictable motion. Avoid asking for things that are not in the picture: the model can move a car, it cannot invent the road behind it convincingly. For a negative prompt, the usual list works: blurry, distorted hands, extra limbs, flickering, watermark, text.
Settings that matter
- Resolution: 480p is four times faster than 720p and often enough for social media; use 720p for a final clip.
- Frames: 81 frames at 16 fps is the standard five-second clip. Longer sequences drift; better to generate two clips and cut.
- Steps: 20 to 30 sampling steps. Below 15 the motion is smeared, above 40 the gain is invisible.
- Guidance (CFG): 4 to 6. Higher values follow the prompt more literally but add flicker.
- Seed: like in Stable Diffusion, the same image, prompt and seed reproduce the same clip; change only the seed to get another motion for the same picture.
A 720p clip takes several minutes on a shared demo GPU, so start at 480p to validate the prompt, then re-run at 720p with the same seed.
Which images animate best
Wan 2.2 works from any photo or AI-generated image, but some pictures are much easier. A clear main subject on a simple background gives clean motion; crowded scenes make the model choose what to move. Landscape framing matches the video aspect ratio, so portrait images are cropped or padded. Sharp, well-lit pictures of at least 1024 pixels on the long side hold detail through the clip; a blurry or heavily compressed source produces a blurry video.
Images generated with Stable Diffusion, SDXL or FLUX animate very well, which makes image-to-video the easiest way to get a short video with a precise look: generate the exact frame you want first, then animate it.
Running Wan 2.2 I2V at home
The weights are on Hugging Face and run in ComfyUI, in Diffusers and in the official Wan repository. The 14B I2V model needs about 80 GB of VRAM in full precision, or a 24 GB card with the fp8 quantised checkpoints and CPU offloading, at the cost of several minutes per clip. The 5B TI2V model generates a 720p, five-second clip on an RTX 4090 in under ten minutes and needs about 24 GB of VRAM. On smaller cards, use the cloud or the free demo on this site.