Step 1: Pick a Reference-to-Video Mode
Text-to-video alone cannot reproduce the action order. The chip, the recovery and the finish have to happen in that sequence, and a prompt cannot hold that timing. Use a mode that accepts an image reference and a video reference together — the image defines who is on the ball, the clip defines what happens. Open the model selector and check for both input slots before adding files. In Dreamina, choose the multi-reference workflow rather than a first-frame-only mode. Make a short draft first, and review the order of the touches before increasing resolution.
Step 2: Upload One Clean Full-Body Photo
Face, hair, outfit and shoes all visible. One person, centered, full body in frame, neutral background. Do not upload the keeper or the defenders: only the player on the ball gets replaced. A cropped portrait produces a cropped character. Use an evenly lit photo with clear edges around the legs, and leave space above the head and below the shoes. Avoid sunglasses, a hidden face, motion blur and extra people. Use your own image or one you have permission to use; a generator accepting an upload does not give you rights to the photograph.
Step 3: Attach the Clip and Paste the Prompt
Cover the whole run. At five seconds the finish gets cut off. Keep the action inside a 9:16 frame so the shot stays in shot. Paste the prompt from the prompt page; the reference clip carries the motion, the prompt locks the identity. Trim idle footage before the approach rather than cutting away the last touch. Check that the ball and goal remain visible after the crop. Set the output length to cover your entire licensed reference, within the model limit. Use Prompt B with the negative lines, then inspect the keeper and defenders before exporting.
Step 4: Fix the Four Common Failures
Change one thing per attempt so you can see what helped. First check the source photo and reference crop, then change the prompt. Keep a copy of the last usable result. If the whole sequence is too complex, test a shorter section without treating that test as the finished video. Only join segments after checking the same face, clothes and ball position at the cut.
- Face drifts mid-run → shorten the clip, restate identity in every segment.
- Ball disappears → name the ball and the net explicitly, not just "scores".
- Feet slide → name the surface (grass) and the run direction.
- Keeper's face changes → block role swapping with the negative lines.
Format, Music and Captions
Export vertically at 9:16, ideally 1080 × 1920, keeping the ball and finish inside the safe area. Add a short caption in your editor after generation so it stays readable. Label the result as an AI recreation. Mute broadcast commentary and music you do not own; use your own audio or music whose licence covers your platform and intended use. Keep generation free of captions, then add your final text in CapCut.
FAQ
How do I make an Ozil AI goal video with my own photo?
Three steps: a full-body photo, a reference clip that runs through the finish, and a prompt that locks identity.
Why does my video change the goalkeeper or the defenders?
The model is re-generating the whole scene. Add negative lines and state "replace only the player on the ball".
Why is the last shot cut off?
Your clip or duration is too short. The finish needs the full run plus the follow-through.
Can I use the original broadcast audio?
No. Use royalty-free music unless you own the track.