You generated a two-shot: woman in red on the left, man in blue on the right. Around second 3 they pass behind a pillar, and when they re-emerge the man is wearing red and the woman is in blue. Or their faces have swapped entirely. The model lost track of which character was which. This is identity-swap, and it is one of the hardest multi-character problems in current AI video.
Fastest fix (as of June 2026): stop relying on text alone. Give each character its own tagged reference image. In Runway Gen-4 References, upload a reference, name it (e.g. woman_red), and prompt with the @ tag: @woman_red walks on the left, @man_blue walks on the right. In Kling, use the Elements feature (1-4 reference images, each bound to a subject). If the swap still happens, generate each character in a separate clip and composite them in post. That last method never swaps because the model only ever sees one person at a time.
Quick diagnosis: which bucket are you in?
| Symptom | Most likely cause | Go to |
|---|---|---|
| Swap happens exactly at occlusion (pillar, hug, crossing) | Weak per-character anchoring | Step 1, Step 2 |
| You can’t tell from the prompt who is who | Prompt doesn’t differentiate characters | Step 1 |
| Reference image has both people overlapping | Fused reference | Step 3 |
| Prompt literally says “switch sides” / “cross” | Action requires a position swap | Step 1 (rewrite) |
| 4s clip is stable, 10s clip swaps | Long duration amplifying weak anchoring | Step 3, Step 5 |
Common causes
Ordered by hit rate.
1. Model has weak per-character anchoring
Text-to-video models reason about a woman in red and a man in blue as a single bag of attributes. Once their pixels mix (occlusion, close framing, hugging), the model is free to reassign which person owns which attribute. This is why even strong 2026 models still swap: the fix is not a better prompt alone, it is giving the model a separate identity anchor per character.
How to spot it: identity swap happens precisely at moments of occlusion or close contact. The model “redraws” both characters when they re-emerge and gets the assignment wrong.
2. Prompt does not strongly differentiate the characters
Two friends walking gives the model nothing to anchor on. A tall woman in a red coat with short black hair, walking alongside a shorter man with curly blond hair in a blue jacket gives much stronger anchors.
How to spot it: re-read your prompt. If you cannot tell from the prompt alone which character is on which side, the model has no reason to keep them straight.
3. Reference image has both characters fused
For image-to-video, your reference image might have both characters in the same crop region or with overlapping silhouettes. The model treats them as one entity.
How to spot it: look at the reference. If you cannot draw a clean bounding box around each character, the model can’t either.
4. Action requires them to swap positions
If your prompt says they switch sides as they walk, the model genuinely did what you asked, but the identities followed the swap. Common in choreographed shots.
How to spot it: does your prompt include switch, pass, cross, or exchange? Position swap can drag identity swap with it.
5. Long clip duration on a multi-character scene
5-second two-shots are usually OK. 10-second two-shots have a much higher chance of identity drift. Same model behavior as single-character drift, but worse because there are two identities to keep apart. Most providers cap a single generation at 5-10s (Runway, Kling, Pika, Sora 2 all sit in this range as of June 2026), so longer scenes are stitched from multiple clips anyway.
How to spot it: generate a 4s and a 10s version of the same prompt. If 4s is stable and 10s swaps, duration is amplifying weak anchoring.
Shortest path to fix
Step 1: Strengthen per-character description in the prompt
This is the cheapest fix and sometimes enough on its own.
# Weak
"Two friends walking through a park."
# Strong
"On the left, a tall woman in a red coat with shoulder-length black hair,
walking next to a shorter man on the right in a blue jacket with curly blond
hair. Maintain positions: woman always left, man always right.
Maintain clothing: red on woman, blue on man, throughout entire clip."
# Use distinct hair color, height, clothing color
# Repeat the assignment at the start and end of the prompt
# Remove any verb that implies a swap: "switch", "cross", "pass", "exchange"Step 2: Use per-character reference tags (best in-tool fix in 2026)
The big change since this article first ran: the leading tools now let you bind a separate reference image to each character, which is far stronger than text alone. Verified current as of June 2026:
# Runway Gen-4 References (supersedes Gen-3 References)
- Upload 1-3 reference images
- Name each one, e.g. "woman_red" and "man_blue" (type a name + Enter to save it)
- Reference them in the prompt with @ tags:
"@woman_red walks on the left, @man_blue walks on the right,
they do not swap positions or clothing"
- Use a clean, well-lit single-subject crop per reference for the strongest anchor
# Kling Elements (Kling 3.0; replaces "Multi-Subject References")
- Add 1-4 reference images, select each subject as an Element
- Front, side, and 45-degree views of a subject improve spatial consistency
- Describe each Element's action; the engine binds them as spatial-temporal anchors
- Works for multi-subject interaction, still drifts on very long clips
# Sora 2 Characters (formerly "Cameos")
- Hard limit of 2 characters per generation
- Each character is a reusable likeness; ~95%+ consistency when both are Characters
- If you need 3+ people, composite (Step 4)
# Pika 2.5 Scene Ingredients
- Upload separate images per character/object
- Model places and animates them without mergingStep 3: Use image-to-video with both characters in the first frame
If your tool has no per-character tags, lock both into a single clean first frame:
# Reference frame checklist
- Both characters clearly visible
- Distinct silhouettes (height, hair, clothing)
- Clean spatial separation, no occlusion
- Strong color contrast in their outfits
# Image-to-video prompt
"Continuation of the depicted scene. Woman in red stays on left throughout.
Man in blue stays on right throughout. No swapping of positions or clothing.
Maintain identities from the reference frame."
# Generate 4 seconds max; longer = higher swap riskStep 4: Generate each character separately and composite
The single most reliable fix, because the model never sees two people at once, so it cannot swap them:
# Generate character A alone in the scene
"A tall woman in a red coat walks through a park from left to right,
empty path, no other people, locked tripod."
# Generate character B alone in the same scene
"A shorter man in a blue jacket walks through a park from left to right,
empty path, no other people, locked tripod, matching lighting and color grade."
# Composite in After Effects or DaVinci Resolve (Fusion)
- Mask / rotoscope each character (or shoot on a plain background for easy keying)
- Layer A on the bottom, B on top
- Adjust timing so they appear in the same shot
- Add a ground shadow under each to anchor themStep 5: Avoid occlusion or rewrite to single-character shots
If swaps persist, restructure the edit so the two characters never share the frame during a risky moment:
# Replace one two-shot with two single-shots
- Shot 1: woman in red walking (4 seconds)
- Shot 2: man in blue walking (4 seconds)
- Cut between them in the editor
# Or generate the two-shot but cut before/after occlusion
- Generate 6 seconds
- Use only the first 3 seconds where no occlusion happens
- Discard the post-occlusion portionHow to confirm it’s fixed
- Scrub the clip frame by frame at the exact occlusion point (the pillar, hug, or crossing). The swap, if any, happens within a few frames of re-emergence.
- Check three things at the last frame: clothing color, hair, and screen position match the first frame for each character.
- Re-roll the same prompt 2-3 times. Identity swap is stochastic, so one clean clip is luck; three clean clips in a row means the fix held.
- If you composited (Step 4), confirm shadows and lighting direction match between the two layers, which is the usual giveaway of a composite.
FAQ
Why do the characters only swap when they pass behind something?
Occlusion is the trigger. While a character is hidden, the model has no pixels to track, so when they re-emerge it re-generates them from the prompt’s attribute pool and can attach “red coat” to the wrong person. Per-character reference tags (Step 2) or compositing (Step 4) give the model a fixed anchor that survives the occlusion.
Which tool is best for keeping two characters straight in 2026?
For a single clip, Runway Gen-4 References (@ tags) and Kling Elements are the strongest because each character gets its own reference image. Sora 2 Characters is excellent but capped at 2 characters per generation. For guaranteed zero swaps regardless of tool, generate each character alone and composite (Step 4).
Can I keep the same characters consistent across multiple separate clips?
Within one clip, current models hit roughly 95%+ consistency with references. Across separate clips it is harder. The reliable approach is to reuse the same named reference images (Gen-4 References persist across sessions once saved) or the same Sora 2 Characters in every generation, then color-match in post.
Does a longer prompt or “do not swap” wording actually help?
Distinct descriptions and an explicit do not swap positions or clothing line help a little, but text alone does not solve identity-swap at occlusion. Treat the prompt as the floor; the real fix is a per-character anchor (reference image) or compositing.
My clip is fine at 4 seconds but swaps at 10 seconds. What do I do?
Generate the scene as two or three short clips of 4-5 seconds each and cut between them, or use per-character references so the anchor holds longer. Duration amplifies weak anchoring, so shortening the generation is often the fastest win.
Prevention
- Default to per-character reference tags (Gen-4 References, Kling Elements) for any shot where identity matters; fall back to single-character composite when it absolutely must not swap.
- Always prompt distinct visual differentiators (color, height, hair) for multi-character scenes.
- Avoid prompts that require occlusion, swaps, or close physical contact between characters.
- Cap multi-character clips at 4-5 seconds; chain in post for longer.
- Build a reference-image library per character (and save them with names) so future shots stay consistent.
Related
- AI video character drift
- AI video face changes mid
- AI video scene inconsistency
- AI video extension loses style
- AI video image to video not following
- AI video product deformation
External references: Runway: Creating with Gen-4 Image References, Kling AI character consistency quickstart.