One photo yields smooth cinematic video. Another produces a jerky clip with a melting face. The difference is almost never in the settings, but in the source image and exactly what you ask it to do.
Three Things That Break Animation Immediately
Small face in the frame. If the face occupies less than one-third of the photo's height, features begin to melt between frames. The larger the face, the more stable the result.
Complex background. When a person blends with the background by color or texture, movement smears. Soft blurred backgrounds, solid walls, and open skies work well. A busy interior or a crowd behind the subject is a risk.
Harsh shadows on the face. Sharp studio lighting or direct flash flicker during motion. Diffused daylight or side window light animate smoothly and predictably.
Which Photo to Choose
The best starting point is a portrait taken on a phone in daylight. Large face, simple background, soft light. The probability of a good result on the first try is maximal.
Looking at the camera or slightly to the side animates better than a profile or a gaze strictly downward. Limbs are better either fully in the frame or cropped along a natural line—waist, shoulders, elbow. A hand cropped in the middle of the palm sometimes reconstructs unexpectedly.
And most importantly: a blurred source yields a blurred video. Nothing here gets fixed during the process.
The Main Prompt Rule: One Action
The most common mistake: "She turns, smiles, hair blows in the wind, the camera pulls back, she takes off her clothes and throws them back, remains completely naked." That's four tasks at once. The neural network will pick one; the rest will be interpreted randomly.
It works like this: one body movement or one camera movement—not both simultaneously.
Concrete is always better than abstract. "Slowly turns head to the right" yields a predictable result. "Natural movement" is a lottery. "Smiles lightly" works stably. "Lively emotional reaction" does not.
The words "slowly" and "smoothly" in the prompt work for the result. "Sharply" and "quickly" increase the risk of artifacts.
Camera Movements: What to Choose
Static camera—all movement is only in the character's body. The most reliable option for portraits, works on any normal source.
Zoom in—the camera slowly moves toward the face. One of the most stable options, creates tension and closeness. A good choice if you don't know where to start.
Pan—the camera rotates horizontally. Works well when the background is simple or blurred.
Orbit—the camera moves around the object. Striking, but requires a high-quality source. On a weak photo, it produces artifacts along the edges.
If the Result Isn't Satisfactory
Regenerating with the same settings makes little sense—the result will be similar. You need to change one variable.
Change the photo. A different angle of the same person often yields a fundamentally different result—especially if the first photo had a complex background or a small face.
Remove everything from the prompt except one action. Don't simplify—just remove. One phrase shows what the model does in its pure form.
Try a static camera. If orbit or pan produce artifacts—remove the camera movement entirely. Often this fixes the image immediately.
Re-crop the source larger on the face and upload again. This changes the result more than any prompt edits.