Confirm that a tool supports multiple faces in video, individual replacement assignments, and leaving people unchanged. Test a segment where people move or cross. Detecting several faces in a still image is not proof that a product will preserve each identity throughout your video.
On this page
Based on official public documentation checked September 15, 2026. Practical troubleshooting steps are editorial suggestions, not results from hands-on product tests.
Separate detection, mapping, and continuity
Detection finds faces. Mapping assigns a reference to a particular person. Continuity means that assignment remains appropriate as the video changes. Before comparing products, decide which of these your project needs.
For a two-person greeting, you may want two different replacement identities. In another scene, you may want to replace only the foreground person and leave everyone else untouched. Both can be described casually as a “group face swap,” but they require different behavior.
FaceFusion documents reference-based and simpler selection modes. This is a concrete example of selection choices in one application, not evidence of how every hosted service tracks people.
Make the requirements explicit
Write a small assignment list before opening a tool: “Person in the blue jacket → Alex’s reference; person in the green shirt → Sam’s reference; background people → unchanged.” Describing the person is more reliable for your review than “left face,” because positions can change.
Ask the following questions of the product interface or its documentation:
- Does multiple-face support apply to uploaded video, or only photographs?
- Can each detected person receive a different reference?
- Can an individual face be left unchanged?
- Can the assignment be checked after a cut or after a person leaves and re-enters?
- Does the supported face count apply to the plan and workflow you are considering?
If a key answer is missing, mark it unconfirmed. Do not infer it from a promotional group image.
Use a difficult but representative test
| Test moment | What you are checking | A reason to pause |
|---|---|---|
| Two people visible together | Both intended assignments appear | The same replacement appears on both |
| People cross positions | The edit follows the intended person | Identities exchange at the crossing |
| One person leaves and returns | Re-entry is handled correctly | The returning face has a different assignment |
| One person turns away | Other faces remain correct | The replacement jumps to a bystander |
| A new shot begins | Mapping is still valid | The next scene is altered unexpectedly |
This is our proposed acceptance checklist. It is not a claim that a specific service passes these tests. Use footage and likenesses with permission, and include the hard moments that will actually appear in the finished project.
Read official feature claims narrowly
Deepswap’s product page advertises simultaneous replacement of up to six faces. That describes a published capacity claim, not a benchmark for difficult crossings or occlusions. Verify the current plan conditions before committing a project.
Magic Hour’s video page links to a multiple-face workflow with individual face mapping. Check that you are in that workflow; a single-face interface should not be assumed to expose the same controls.
Our tool comparison puts these products in the context of browser workflows, mobile apps, and larger creation platforms. It does not rank their output quality.
Plan the review and the cost together
Count tests as part of the project, not just the final export. Keep a log of the reference assignments and the segment used. If a tool charges for each render, repeated experiments can matter more than the duration of one finished clip. Read how it counts video credits rather than assuming every face has the same billing effect.
Splitting a video into shots can make assignment easier to review. It also creates an editing task: check audio continuity, cut points, and the transition between shots after recombining them. Use the free-tools guide to check export restrictions before investing in that workflow.
Common questions
Does support for six faces mean six faces will always look good?
No. A published capacity is not a quality guarantee. Inspect your own representative clip before relying on it for delivery.
What if the mapping looks right but the face jumps?
Review the frames around the jump using the flicker checklist. Selection and visual blending can require different experiments.
Sources
- Deepswap: published multiple-face capacity.
- Magic Hour: video and multiple-face workflow.
- FaceFusion: Face Selector — an example of explicit selection modes.