You made the zombie trend video. The ash, the hug, the hard cut to the beach. And the person in it looks like your cousin. Or like you in shot one and a stranger in shot three.
This is the most common complaint about the trend, and it's rarely bad luck. AI video models lose a face in a few predictable ways, and each one has a fix. This post covers the why, then the fixes in the order that saves you the most credits.
Short version (the checklist):
- One person per photo, facing the camera, even daylight, no filter
- Face fills at least a third of the photo height after cropping
- Make a still image first and check the likeness before making any video
- Paste the same description of each person into every prompt, word for word
- Same clothes and one "anchor" object (ring, scarf, collar) in every shot
- Keep each shot short, around 4 seconds
- Turn gently: grey skin and clouded eyes, not "rotting"
- Judge the likeness in the warm ending, not the zombie shots
1. Why the face changes in the first place
An image-to-video model gets one flat photo of you. Everything else, your profile, your teeth when you laugh, how your face moves, it has to invent. Four things make that invention go wrong:
- Every shot starts from scratch. If you generate four shots separately, the model rebuilds your face four times. Small guesses add up, and by the last shot you've drifted.
- Length makes it worse. Within one clip, the further a frame is from the start frame, the less it's tied to your photo. A 10-second shot drifts more than a 4-second one.
- The zombie look removes the clues. Skin tone, eye colour and lip colour are some of the strongest signals of who someone is. Grey skin and white eyes take them away, so the model leans on its own idea of "a zombie", not on you.
- Two people share one prompt. When both people are in the same frame, details can leak between them: his beard shows up on her, her hair colour on him.
Top row: the same man, regenerated shot by shot, slowly becomes someone else. Bottom row: same man, locked with a reference still and a fixed description. Image created with AI.
2. Fix the photo first
Most likeness problems are decided before you type a prompt. The model can only keep what the photo shows.
- Front-facing, eyes open. A profile or a look-away hides half the features the model needs.
- Even light. Window light or open shade. Hard sun or a ring light flattens or splits the face.
- No beauty filter, no heavy edit. The model will copy the filter's smoothed skin and bigger eyes, so the result looks like the filter.
- Big enough. Crop so your face takes up at least a third of the photo height. A tiny face in a group shot gets upscaled into a guess.
- Glasses only if you always wear them. If they're part of how people know you, keep them, and mention them in every prompt.
- Recent. A photo from five years ago gives you a video of five years ago.
For two people, use two separate photos that look different: different backgrounds, different clothes. Two near-identical selfies are how faces start to blend.
3. Test the likeness with one still before any video
This is the step that saves the most money. Before you spend video credits, generate a single still image of the warm ending: the two of you alive, in good light, using an image model that accepts reference photos.
If that still doesn't look like you, the video won't either. Fix the photo or the description now, while each try is cheap. When the still is right, it becomes the start frame for the video, and the model has a much better anchor than your original photo.
4. Write each person's description once and reuse it exactly
Models don't know names, so they rely on what you describe. If shot one says "dark curly hair" and shot three says "brown wavy hair", you've asked for two people.
Write one identity block per person and paste it into every prompt without changing a word:
Person A: woman about forty, curly dark hair streaked with grey,
shoulder length, sage-green linen shirt with rolled sleeves,
thin silver necklace, stays human.
Person B: man about forty, short black hair, trimmed beard,
navy wool coat, wedding ring on left hand, turned.
Then build each shot prompt as: [identity blocks] + [what happens in this shot] + [light and camera]. Only the middle part changes.
Same woman, three ways. The hair, shirt and necklace never change, so even the turned version reads as her. Image created with AI.
Keep the clothes the same in all four shots. Outfit changes are where faces drift most, because the model treats a new outfit as permission to make a new person. Pick one anchor object (a ring, a scarf, a dog's collar) and name it in every prompt. When a face shifts slightly, the anchor still tells the viewer it's the same person.
5. Keep shots short and start each from a still
- Make each shot about 4 seconds and join them in the edit. Four short shots stay closer to your face than one 15-second shot.
- Start every zombie shot from the same cold still, and the ending from the warm still. Don't generate video straight from text, because text-only shots invent new faces.
- Make two or three takes of each shot and keep the best. Regenerate only the shot that's off, not the whole video.
6. Turn them gently
The more extreme the zombie, the less of the person survives. "Rotting flesh, exposed bone, sunken face" asks the model to rebuild the face from the skull up.
What keeps the likeness:
Same face, same hair and same clothes as the reference.
Pale grey skin, clouded white eyes, dark circles,
torn sleeve, calm expression. No blood, no wounds.
This also keeps you under most models' content filters, which often block gore and bites.
7. When the two faces blend together
- Use separate photos with different backgrounds and clothes (see section 2).
- Say where each person is in the frame, in every shot: "Person A on the left, Person B on the right."
- Make the hug from a still where both people are already correct, not from text.
- If the wrong person turned, swap which photo is A and which is B. Don't try to fix it in the prompt.
8. Judge the right shot
The zombie shots are supposed to hide features. Friends recognise you in those shots from your hair, clothes and the anchor object, not your face. Judge the likeness in the warm ending, where your face is fully visible. If the ending is right and the zombie shots are close, the video works.
Pets are easier than you think
The same rules work for dogs and cats, with one change: their markings do the job a face does for people. Shoot at their eye level, without flash (it changes eye colour), and make sure the patches, the ear shape and the collar are in the photo. Name the markings in the identity block: "black and tan, white blaze on chest, left ear folds over, red collar with a bone-shaped tag."
Disclosure: I build AI Zombie, a site that makes this video from two photos. It does the stills, the four beats and the cut for you, so most of the steps above happen inside it. It's paid (one-time credits from $6.90, no subscription), and photos are deleted after 7 days. If you'd rather do it yourself, the free photo guide goes deeper on section 2.
Got a face that keeps drifting no matter what? Leave a comment with what you've tried, and I'll suggest a fix.
Top comments (1)
tr.ee/dev-to