Wan 3.0 image-to-video: first frame, optional last frame

Image-to-video on Wan 3.0 starts from a still you upload. The first frame is required. A last frame is optional, and the model is asked to land on it. The prompt should describe the motion and the sound, not re-explain a picture the model can already see.

What the upload will accept

The form checks the file before it spends credits:

• JPG, PNG, WebP, or BMP.
• 20MB or smaller. PNG and BMP are redrawn as JPEG when they need a resize.
• The short side must be at least 240 pixels.
• The long side cannot be more than 8 times the short side. A very thin banner is rejected.
• If the long side is over 8000 pixels, the image is scaled down.

A blurry crop of a face, or a collage of five people, is accepted by those rules and then fails as a video. Use one clear subject if you want that subject to survive the motion.

First frame and last frame

The first frame is the opening picture. Treat it as locked: wardrobe, time of day, and where the person is standing are already decided. Use the prompt for what moves. “She turns her head toward the window and the curtain lifts once” is a motion note. “A woman in a red coat stands by a window” fights the still if the coat in the photo is blue.

The last frame is the picture the clip should end on. Use it when you know the end and not the middle: a closed box that should finish open, a product that should finish facing the other way, two poses of the same person. Skip it when you only know the start. An unrelated last frame makes the middle a scramble between two photos.

Switching into Image to Video from the default 16:9 text setting moves the aspect control to adaptive, which lets the still suggest the frame. You can still force 9:16 or 1:1 if the delivery format matters more than the photo’s crop. Forcing a tall frame on a wide photo will crop or pad; the resize control on the provider side is not a design tool you get to art-direct here. If the frame must be exact, crop the still before you upload it.

A prompt that matches a still

Keep the same four beats as text mode, but start from the photo. One workable shape:

• 0–2s: hold the opening, with a small motion already in the photo if the still is sharp enough (steam, a curtain, hair).
• 2–6s: one camera move and one subject move. A slow push, and a hand lifting a cup.
• 6–8s: the end pose. If you uploaded a last frame, say that the clip should arrive on that picture.
• Audio: the sounds that belong in that room. A cup, a street, no music, or a single spoken line.

Duration, resolution, and credits use the same table as text-to-video. An 8-second 720P test is 4 credits. Do that before a 1080P version of a still you have not seen move yet.

When a still is the wrong input

A first frame is a bad fit when the person must walk into a new room, change clothes, or be seen from the back and then the front with the opening photo showing only a face. The model will try, and the face often drifts. For a photo that should stay recognizable while the body moves through a preset scene, use a template. Those send the photo as a reference, not as a frozen first frame. The reference guide spells out that difference.

Safety checks run on the upload. A blocked image is rejected before the video job. Do not upload anyone you do not have permission to film, and do not upload a minor. That rule is the same on the templates.