Wan 2.1 image to video takes a single still frame and generates a short clip that continues from it. You upload a picture, describe the motion you want, and the model animates forward from that starting point. Everything runs on hosted hardware, so the only thing your own machine has to do is load a web page.
This page is about the 2.1 generation specifically. If you want the newer releases, /wan-2-2 and /wan-2-5 cover those, and /wan-2-1 is the general hub for this generation. The broader image-driven workflow lives at /image-to-video.
The model receives your image as the first frame and predicts what happens next. It is not a slideshow tool that pans and zooms a static picture. It synthesises new frames, which means subjects can move, hair and fabric can shift, and light can change across the clip. The original image stays as the anchor, so the opening frame looks like what you uploaded.
Because the image carries so much of the information, your prompt has a different job here than it does in /text-to-video. You are not describing what should exist. You are describing what should happen.
Start with a clean source image. Compression artefacts, heavy noise, and aggressive sharpening all get carried forward and amplified once the frames start moving. If your file is oversized or in an awkward format, the image compressor and image converter handle that before you upload.
Compose for motion. A frame where the subject is centred with some space around them gives the model room to work. A tightly cropped face with no background context leaves very little to animate.
Write the prompt as a sequence of events, not a list of nouns. "She turns her head slowly to the left, hair drifting, background lights pulsing" gives the model a trajectory. "Woman, beautiful, cinematic, 4k" gives it nothing to move.
Keep the motion modest. Small, believable movement reads as real. Large movement across a short clip tends to smear or warp, especially around hands, faces, and thin edges.
Treat the first render as a test. Watch it once for the overall motion, then again for artefacts. Most of the improvement comes from the second and third attempt, not the first.
The prompt is the biggest lever, but not the only one.
Aspect ratio should match your source image. Forcing a portrait photo into a landscape frame means the model has to invent the missing regions, and those invented areas are usually the weakest part of the clip.
Clip length trades against coherence. Longer outputs give the motion more time to develop, but drift and identity loss accumulate the further you get from the source frame. Short clips stay closer to the original.
Resolution affects how much fine detail survives. Higher settings preserve texture but cost more to render and can expose flaws in a soft source image.
Seed control matters when you want to iterate. Reusing a seed while changing one variable lets you see what that variable actually did, instead of guessing against a completely different roll.
The credit cost of a run is displayed next to the form before you submit, so you can see what a configuration costs before committing to it. Current plans are on /pricing.
This is not a tool for reconstructing a scene you never photographed. If you have no starting image, /text-to-image or /ai-generator is the right entry point.
It is not built for long-form video. Clips are short by design, and stitching many of them together is a separate problem that /video-extend addresses.
It will not reliably produce readable text, precise hand gestures, or exact physical interactions. Those remain weak spots.
It is not a face-swap or character-consistency tool. If your goal is driving a specific character with a reference performance, /wan-animate and /wan-2-2-animate are closer to what you want.
And it is not a local install. This site runs the model for you in the browser. We do not distribute model weights and we offer no downloads of any kind. If you are looking for files to run on your own GPU, this is not that.
Do I need a GPU or any software installed?
No. The rendering happens on hosted hardware. You need a browser and an image.
Can I use a photo of a real person?
You are responsible for having the rights to whatever you upload. Do not submit images of people who have not consented.
Why did my clip drift away from the original image?
Usually the motion you asked for was too large for the clip length, or the source image had little detail for the model to hold onto. Shorten the motion, or shorten the clip.
Can I download the model weights here?
No. We do not distribute weights and there are no downloads on this site. We run Wan models for you and return the rendered video.
What if I only have text and no image?
Start at /text-to-video, or generate a still first and bring it back here.