How to fix faces and hands in AI video

This is a practical repair guide, not a model-quality ranking. The examples are proposed production plans, not before/after tests we performed. Magic Hour publishes the guide and links its own workflows. Official references checked October 2, 2026; a model’s controls and a host’s available settings can differ.

Choose the repair by the actual defect

Treat these as editorial triage choices, not guaranteed fixes. If exposure or texture pulses while anatomy remains correct, use the AI video flicker guide. If the same person differs between separate clips, use the character-consistency guide.

1. Find the first broken frame in the downloaded original

Keep an untouched copy of the downloaded clip. Watch it at normal speed, then step through the time range where the problem appears. Record the affected region and first visible change: for example, “right thumb merges with the bottle from 2.4 to 2.9 seconds.” Compare against the original source rather than memory.

For a face, compare identity, expression and visible structure without mistaking a shadow or natural turn for a different person. For a hand, inspect finger count where visible, joint shape, grip, contact and the object being held. A partly hidden finger is not automatically missing; follow it through the surrounding frames before deciding.

Check whether the defect exists in the original or only after editing, export or upload. If only one player shows stutter, compare another player and the local file before paying to regenerate. A delivery problem and an anatomy error need different remedies.

2. Correct the source before writing a longer prompt

Inspect the starting image at the intended crop. Reject a source with an already fused hand, distorted eye, blurred face or impossible grip. Choose a permitted reference that actually shows the intended person and action. Extra descriptive words do not establish that an incorrect source will be repaired.

Runway’s image-to-video prompting guide says the input image supplies composition, lighting and style, and warns that blurry hands or faces can become more pronounced in animation. It recommends starting with essential motion and refining incrementally. That is documented guidance for its generation workflow, not a universal architecture explanation or quality guarantee.

A starting frame, a character reference and a pose reference are different controls. Use the input type supported by the selected mode. Do not assume that uploading additional photographs trains an identity or that every host exposes the provider’s full reference system.

3. Make the difficult action easier to evaluate

As a troubleshooting experiment, isolate the action that fails. Compare a small head turn with a fixed camera before combining a turn, orbit, changing light and a moving foreground object. Keep other inputs constant so the next attempt tells you whether that change helped.

For hands, make the intended contact clear in the reference. A hand holding a bottle and a hand transferring it to another person are different briefs. If the transfer repeatedly fails, consider one shot showing the first person holding it and another showing the receiver holding it. Check the join; splitting the action does not automatically preserve continuity.

Frame important details so you can assess them in the delivered video. Cropping a hand out can be a legitimate editorial choice when the action does not require it. It is insufficient when the viewer needs to see how a product is used. Keep the required demonstration intact instead of hiding its essential step.

Two suggested motion requests

These requests are starting points for a controlled attempt, not tested recipes. Use a clean reference that already shows the required face or hand. Inspect the actual result and confirm that your selected mode accepts the source and duration.

A preservation instruction is a requested outcome, not a hard constraint. Reject the result if identity, anatomy or the product changes despite the wording. For broader prompt diagnosis, see AI video prompt troubleshooting.

4. Replace the smallest usable shot, not an arbitrary frame

When the failure occupies part of a sequence, retain accepted footage and plan a replacement shot around the broken action. Use an input and duration the chosen tool actually supports. A five-second generation setting is different from a timeline trim or an arbitrary fractional-second repair.

Compare the replacement at both joins. Pose, lighting, eye line, grip and object position should fit the preceding and following footage. If the replacement is anatomically correct but creates a visible jump, keep working on the edit or redesign the transition. Check the full motion, not only the replacement’s first frame.

5. Choose a tracked patch or a generative edit deliberately

A conventional compositing repair uses a valid replacement region, a mask and tracking to keep it aligned over time. It is worth evaluating when the surrounding shot is usable and the patch can match perspective, lighting and occlusion. A corrected still pasted over several frames may float, slide or cover the wrong object.

Adobe’s Premiere mask-tracking instructions explain how a mask follows a moving subject, then require review and mask or keyframe adjustment. Tracking aligns a selected region; it does not itself reconstruct correct fingers or establish a matching identity. Inspect edges and every moment another object crosses the patch.

A generative edit produces a new version of existing footage. Magic Hour’s AI Video Editor documentation describes source video, a text edit and a selected time range. For that route, see how to edit a video with text prompts. Recheck the entire result for unrequested changes; a render marked complete is not a repair verdict.

Runway’s Edit Studio guide documents editing a selected keyframe and applying that change to video, optionally within a selected range. This is different from repairing one still and assuming the neighboring frames follow. The source explains controls, not a guarantee that every hand or face can be recovered.

Three illustrative repair plans

Portrait loses identity in profile: identify the first change during the turn. Try a source that clearly shows the required angle and a smaller action. Review the point where the face becomes partly hidden. If the approved identity still changes, use a different shot or workflow rather than approving a convincing front-facing frame.

Fingers merge during a product handoff: inspect the hand, contact and package together. Test a simpler hold or separate donor and receiver shots. If exact handling is essential to the demonstration, recorded footage may be the stronger production choice. A smooth-looking transfer with an impossible grip still fails the brief.

Crowd faces look wrong after a crop: compare the original and final crop. Decide which people must be recognizable. Reframe or regenerate around the important subject when practical. Do not present altered background people as faithful documentary evidence, and do not claim an upscale recovered an identity the original did not contain.

Why upscaling or a face swap may not solve the problem

An upscale changes the delivered pixel dimensions; that alone does not prove correct anatomy or identity. Compare the result with the acceptance rule. Sharper edges around a fused finger still represent the same failed action.

A face swap targets facial identity, not a malformed hand or the whole body. It may be worth evaluating for a consented identity-replacement task, but review expression, head turns, occlusion and boundaries. Choose the operation you actually need instead of assuming every face-related tool is a general anatomy repair.

A final acceptance checklist

Watch the full exported sequence with sound at its intended size. Review the repaired interval frame by frame. Confirm identity where required, plausible visible anatomy, correct grip and contact, stable product details, believable edges and clean joins. Check the uploaded version as a separate delivery step.

Keep the original, accepted output, source, settings, attempts and rejection notes. Count failed attempts and repair time when budgeting; cost per accepted AI video gives a reusable calculation. Use the AI video QA checklist for the whole deliverable, not only its anatomy.

For a different generation route, use the AI video generators guide. The Image-to-Video documentation explains generating motion from a still. These links identify workflows; this guide does not establish one universally best model or a measured repair success rate.