FlyAIgh
Home/Blog/Guide

How to Replace Two Characters in a Video Without Losing Track of Who Is Who

Published October 28, 20268 min read

Two references are only the start. Follow each identity through position changes, glasses and a drum close-up, using a real APT template result that still contains mistakes.

Choose face swap or whole-character replacement

To replace two characters in a video, assign a separate reference image to each source identity. Keep those assignments when the characters change position or disappear behind props, then check both identities after every cut.

If you only want new faces, look for a multiple face swap workflow with individual face mapping. If hair, body shape and clothing should follow new character designs, you need whole-character replacement. This guide demonstrates the latter with FlyAIgh’s built-in APT duo template; it is not a face-only retouching tutorial.

Watch the complete example, including the missed swap

The source is a 14-second pink-studio performance by two anthropomorphic cats. The result below was generated with WAN 3.0 at 720p using two character references. Most scenes show silver-haired humanoid characters, but the drum shot still shows the original orange cat. Watch through to the end before judging it.

Source: FlyAIgh’s 14-second APT-style cat template. Presented without audio.
Historical replacement result: full 14-second take, 720p, without audio. The drummer remains a cat; this is a review example, not a fully successful swap.

Cat ears and pointed teeth also remain on the humanoid characters. A task marked completed proves a file was generated; it does not prove both roles were replaced correctly. This example is useful precisely because a convincing opening would miss the later failure.

Assign references by identity, not by position

Give the two source identities stable names before uploading anything. Here A means the cream long-haired ragdoll with dark face markings, and B means the orange tabby. Left and right are temporary positions. The tabby’s cap is another temporary cue: it is absent in the drum shot.

Use a separate image for each replacement character, rather than a group portrait. Show the face clearly and include the body and outfit when they matter. If the workspace offers a cleaned character sheet, inspect it before selecting it: newly generated views can introduce details that were not in your original design.

Source identityWhere your reference goes
A · Cream ragdoll, dark face and earsRagdoll role slot → your first character
B · Orange tabby, including the drummer without a capTabby role slot → your second character

Make your version in the APT workspace

  1. Open the APT duo template and play its sample. Confirm that its action and framing suit the clip you want.
  2. For each role you want to replace, upload or select that character’s image in the matching role card. To keep one original cat, leave its optional slot unchanged. At least one role must be replaced.
  3. Review the reference check. Keep the original image or inspect the optional extracted sheet before using it. Image preparation may have a separate quote.
  4. Choose an available video model, resolution and sound setting. Open the prompt preview and verify which image is assigned to each source identity. The historical take here used WAN 3.0 and 720p.
  5. Read the current quote, then generate when the inputs are ready. Compare the complete result with the sample; inspect the retained role too if you only replaced one.

Open the APT duo template

Describe the exceptions, then check the prompt

The old request identified the tabby by its cap and sunglasses. The current template also states that the drummer without a cap is that same character, and asks the model not to inherit unwanted cat features. That clarification addresses ambiguity; this case does not isolate it as a proven fix.

Replace the cream ragdoll with character A and the orange tabby with character B. The orange drummer without a cap is also B. Keep these assignments when their screen positions change. Follow each reference’s appearance and clothing, without mixing the two designs. Keep the source action, cuts and pink studio; do not carry over cat features that are absent from the references.

Read the actual request for this take (Chinese original)

The original text is preserved in every language. It contains the requested actions and constraints, which the result did not fully satisfy. It predates the explicit no-cap drummer and unwanted-cat-feature instructions.

镜头运动、剪辑节奏、光影和粉色影棚完全按照 视频1,每一个镜头的景别、机位和切换时间点都和 视频1 一致。把 视频1 里奶油白长毛、脸和耳朵是深棕色的布偶猫换成 图1 的角色,长相、体型、发型或毛色和服装都以 图1 为准。把 视频1 里反戴黑色棒球帽、戴墨镜的橘色虎斑猫换成 图2 的角色,长相、体型、发型或毛色和服装都以 图2 为准。每个角色的动作、表情、口型和节奏都完全模仿 视频1 里对应的那只猫:一起跳舞,面对面坐在地上拍手划拳,戴着粉色星星形状的眼镜凑近镜头,仰头,躺在地上,靠在一起,打鼓,最后并肩站着。两个角色的长相不能互换,也不能混在一起。画面里只有这两个角色,被换掉的猫不能再出现,不要多余的人或动物。粉色背景、音箱、鼓和道具保持 视频1 的样子。无背景音乐,只有环境音。不要字幕,不要文字。

This is an editable explanation of the mapping, not the exact request used for the video above. In the APT workspace the role cards already supply the assignments: use the optional note only for extra requirements, and inspect the generated prompt preview. Do not paste invented image-number tokens into the note.

Review every cut before choosing a take

Use these approximate template timeline ranges as navigation, not exact edit boundaries. There are cuts within some ranges. Pause just after each cut and when a hidden face becomes visible again.

TimelineWhat to inspectDecision to make
0–3 sDance and close-upsDo both appearances match their assigned references? Check ears, teeth, hair and outfit.
3–6.1 sSeated hand gameFollow each identity as framing changes; inspect touching hands and body overlap.
6.1–7.8 sStar-shaped glassesCheck identity before and after the glasses cover the face, not only the prop itself.
7.8–11.5 sLeaning and floor movementInspect reappearing faces, limbs, clothing and any identity exchange.
11.5–14 sDrums and closing duoThe drummer should also be replaced. In this take it is not: reject it if every appearance must change.

Keep a take only if it meets your intended use across the whole clip. A useful reaction shot may survive a failed full take. Cutting around a missed swap can remove that moment, but it also removes the action; it cannot repair the missing replacement. If the drum performance is essential, revise the relevant identity description or reference and generate another attempt, then review again.

Case record and further reading

The replacement take comes from our 2026-10-09 APT test: WAN 3.0, 720p, 14 seconds, two replacement references. The web copies retain the complete sequences and remove audio. It is one historical result, not a model ranking or a success-rate measurement. The template is our cat remake inspired by ROSÉ & Bruno Mars’s APT.; it is not an official collaboration.

Magic Hour’s help distinguishes individual face mappings from whole-character replacement. Runway’s reference guide provides background on preparing character images, but its settings are not FlyAIgh’s contract. The actual workflow and example above come from our own template, request record and footage.

Single-person replacement guide

Make a character perform an existing dance

Character consistency across shots

Frequently Asked Questions

Can I replace two people with two different characters?

The APT template has separate role slots, so each source character can receive a different reference. Check the mapping and every shot afterward; two uploaded images do not guarantee distinct, stable identities throughout.

Can I change only one person and keep the other?

Yes, in APT you can leave one optional role as the original and replace the other. At least one must change. This is a generation instruction, not a guarantee that the retained character stays pixel-identical.

Does multiple face swap also change the clothes?

Face swapping targets the face; whole-character replacement can request new hair, body and clothing. Choose by the change you need. This APT example uses whole-character replacement and still contains visual errors.

Why did the replacement disappear after a cut?

In this take the capless drummer remained the original cat. A description based on a removable accessory leaves ambiguity, but the video alone cannot prove the model’s internal reason. Identify that shot explicitly and check the next attempt.

Should I upload a photo with both characters together?

For APT, prepare one identity per role image. A group picture makes the intended identity less explicit. Inspect any automatically prepared reference sheet for changes to the face, outfit or body before using it.

Can I use these two role slots with my own uploaded duo video?

This guide covers the built-in APT template. Its two named role slots are not automatically available in the separate own-video swap workspace. Check that workspace’s current inputs and model controls before planning a custom two-person edit.

Is generation free, and can I keep the original music?

FlyAIgh video generation requires a paid account and credits; use the current quote, including any optional image preparation. These examples are silent. Use source footage, identities and audio you have permission to use; replacing characters does not grant music rights.

APT. · Two in the pink room

Dance, a hand game on the floor, star shades, the drums

Use this template