Applies to: Users animating an uploaded or previously generated image. I2V means image-to-video.
Use this guide to animate an image into video while keeping the subject accurate, and to animate photos of real people without losing their likeness or hitting content blocks.
General rules for image-to-video prompts:
Describe the motion, not the image. The video model can already see your image. Re-describing it wastes prompt space and can cause the subject to drift. Your video prompt should answer: what moves, and how does the camera move?
Fix problems in the image first. A strong starting image with a short motion prompt beats a weak image with a long prompt.
The video's aspect ratio usually comes from the input image.
Avoid wording that implies the subject's appearance changes, unless that's what you want.
Example image-to-video prompt: "She turns toward the camera, hair moves in the wind, subtle smile, gentle push-in."
Animating photos of real people (faces):
Faces are the hardest thing for video models to preserve. Follow these six rules for the most reliable results with real faces:
Start with the best possible source image: sharp, high resolution, evenly lit, face clearly visible. Nothing covering the face — no sunglasses, hands, or hair across it. If your photo is soft or low resolution, upscale it in Venice first, then animate the upscaled version.
One face per image. Multiple faces multiply errors, and background faces break first.
The face should fill a meaningful portion of the frame. A medium close-up works best. If the person is small in a wide shot, the video model has too few pixels to preserve their likeness — crop closer before animating.
Front-facing or three-quarter angles with a neutral or mild expression animate most reliably. Extreme angles and extreme expressions distort faster.
Keep the motion subtle. A head turn, a blink, a smile, a slow push-in — these preserve identity. Spinning, running, or fast head movement forces the video model to re-invent the face every frame, and the likeness falls apart.
Do not describe the person's appearance in the video prompt. Text descriptions compete with the image and pull the face away from the real person. Motion and camera only.
Avoiding content blocks when animating real people:
Do not name celebrities or public figures in the video prompt. Naming a real person is a common block trigger even when the request is harmless. Let the image carry the identity and describe only the action: "The man smiles and nods" — not "Make [celebrity name] smile."
Remove identity-flagging words if blocked: names, "famous," "celebrity," "lookalike." Re-submit with motion-only phrasing.
Some blocks cannot be rephrased around. Venice does not support creating NSFW (not-safe-for-work, meaning sexual or explicit) content of real people from uploaded photos. These requests are blocked regardless of wording — this is a policy limit, not a prompt problem.