How to Fix Face Drift in AI Duo Videos

Face drift — the moment a subject stops looking like the person in your photo — is the most common quality complaint in two-photo AI videos. You cannot eliminate it completely with current models, but you can push the odds firmly in your favor.

Why does face drift happen?

The model is not copying your photo pixel by pixel. It reconstructs a moving person that matches the photo's identity, then animates it through a 15-second performance. Every frame is a fresh decision, and over hundreds of frames the identity can slowly wander — especially during fast motion, hand gestures near the face, or dramatic lighting changes.

Knowing this explains every fix below: anything that makes the identity easier to read reduces drift.

Fix 1: Start with a sharper photo

This is the single biggest lever. A face that is small, blurry, or half in shadow gives the model very little identity signal to hold onto. Use a photo where the face is large, lit evenly, and looking at or near the camera. The full checklist is in the photo selection guide.

Fix 2: Keep gestures away from the face

Templates with big arm swings, hair flips, or hands near the face are the hardest for identity. If your result drifts, try a calmer template — stillness is not a flaw in a 15-second performance; it reads as confidence on screen.

Fix 3: Swap the photo order

The two photos are not symmetric: the left photo drives the left performer and the right photo drives the right one, and each template's motion was built around its original staging. If one subject looks great and the other drifts, swap the two photos and generate again — the drift often follows the harder template position, not the person.

Fix 4: Draft at 480p, then re-export

Never debug a face at 45 credits. Generate the test at 480p (20 credits), inspect both faces frame by frame, and only spend 720p or 1080p on the version that already looks right. The quality comparison breaks down when each tier is worth it. Failed generations return credits automatically, so a bad draft costs you nothing but a minute.

What is a realistic expectation?

With clean photos and a calm template, most generations today hold identity well for the full 15 seconds. Occasional softening during the busiest seconds is normal and usually invisible at social-media viewing sizes. If a specific generation is unusable, regenerate — variance between attempts is real, and the second or third try often lands.

Ready for another take? Open the generator.

Frequently asked questions

Does higher resolution reduce face drift? No. Drift is an identity problem, not a pixel problem. 1080p makes a drifting face sharper, not more accurate.

Will my video always fail if the photo is low quality? Not fail — wander. Low-quality photos usually still generate, but identity stability drops noticeably.

Is drift a bug I should report? If one attempt drifts badly, simply regenerate. If every attempt with a clear photo drifts severely, let us know through support with the task ID.