Applies to: Users generating video from an uploaded image or reference (where supported).
Use this guide to diagnose why a generated video drifted away from your reference image or lost a real person's likeness — the most common image-to-video complaint. It almost always traces to one of these seven causes:
You re-described the person in the video prompt. The video model already sees your image. Describing the subject's appearance in text creates competing instructions — the video model blends your words with the image and drifts away from the real person. Describe only the motion and camera: "She turns toward the camera, gentle push-in" — nothing about her face, hair color, or build.
The reference image is too weak. The image must carry enough information to anchor the video model: clear, well-lit, high resolution, face unobstructed (no sunglasses, hands, or hair across the face), subject prominent in frame. If the photo is soft or low resolution, upscale it in Venice first, then animate the upscaled version — the video model cannot preserve detail that isn't there.
The video prompt implies transformation. Wording like "she becomes," "transforms," "now wearing" tells the video model to change the subject. Remove it unless intended.
Too much motion, too fast. Large or rapid movements (spinning, running toward camera) force the video model to re-invent the face frame by frame. Start with subtle motion — a head turn, a smile, a slow push-in — and identity holds far better.
The face is too small in frame. If the person occupies a small portion of the image, the video model has too few pixels to preserve their likeness. Crop closer or use an image where the subject fills more of the frame. A medium close-up is the sweet spot.
Extreme angles or expressions in the source photo. Front-facing or three-quarter angles with a neutral or mild expression animate most reliably; extreme poses distort faster.
For multi-clip consistency: reuse the same reference image and identical wording across clips; drift typically shows up by the third or fourth clip when references are generic.
If the video generation is blocked instead of wrong: naming a real person or celebrity in the video prompt is a common block trigger. Remove names and words like "famous" or "lookalike," let the image carry the identity, and describe only the action.
Note: Venice does not support creating NSFW (not-safe-for-work, meaning sexual or explicit) content of specific people from uploaded images — such requests will be blocked regardless of prompt phrasing.