Learn how to fuse multiple reference images into one consistent character for AI video, with prompt structure, keyframe control, and a repeatable workflow.
You generate one portrait you love. The face is right, the light is right, the mood is right. Then you cut to a second shot, and the character has quietly become someone else: a different nose, a different jawline, a different age. Nothing breaks the illusion of film faster than a face that changes between cuts.
The fix is not a better single image. It is a better set of images, combined deliberately. When you fuse several reference frames into one coherent identity and then hold that identity steady across shots, AI video stops looking like a slot machine and starts looking like a production. This guide covers how image fusion works in practice, how to build a reference set that survives motion, and how to run the workflow end to end.
Why one perfect portrait is never enough
A single reference image encodes a lot of information that has nothing to do with your character: one camera angle, one light direction, one expression, one lens. The model faithfully learns all of it, including the parts you did not want.
That is why a beautiful three-quarter portrait often produces a video where the character can only ever exist in three-quarter view. Ask for a profile and the model invents a new nose. Ask for a wide shot and the face drifts. Ask for a different time of day and the skin tone shifts with the lighting it memorised.
The problem compounds across a sequence. Each shot is generated with a slightly different interpretation of the prompt, and small deviations accumulate. By shot four, you have a sibling rather than the same person. Audiences may not name the error, but they feel it as cheapness — the same feeling as a continuity error in a low-budget film.
Consistency, not raw generation quality, is what separates watchable AI video from forgettable AI video. And consistency is a data problem before it is a model problem: what you give the model to anchor on determines how stable the result can be.
What image fusion actually does
Fusion is the process of combining several inputs so the model extracts a shared identity rather than copying one frame. There are two broad approaches, and knowing which one you are using changes how you prepare your inputs.
Averaging versus semantic fusion
Naive averaging blends pixels. It works for texture and colour but destroys faces, because faces are not aligned across angles. Blend a front view and a profile and you get a smear.
Semantic fusion is different. The system encodes each image into an identity representation — the features that persist regardless of pose, expression, or lighting — and then merges those representations. The output is a stable identity vector you can reuse, rather than a merged picture you have to re-upload.
In practice, most modern tools sit somewhere between the two. They extract facial structure from several images, weight the cleaner ones more heavily, and discard noisy detail. Your job is to feed them inputs that agree with each other.
The three inputs that matter most
Across tools, three things dominate the final result:
- Facial geometry. Jawline, eye spacing, nose bridge, brow. These come best from neutral, evenly lit, front-facing shots.
- Surface and colour. Skin tone, hair colour, freckles, scars, tattoos. These come best from sharp, close, natural-light images without heavy grade.
- Silhouette and wardrobe. Hair length, body shape, coat, hat, uniform. These come from wider shots.
If you only supply close-ups, the model learns a face but not a person. If you only supply wide shots, it learns a costume but not a face. Fusion works when the inputs cover all three layers.
Building a reference sheet that survives motion
Before generating anything, assemble a small, opinionated reference set. Six to twelve images is usually the sweet spot; beyond that, marginal images add noise rather than signal.
Cover angles, not quantity
Prioritise coverage over volume:
- One clean frontal portrait, neutral expression
- One three-quarter left, one three-quarter right
- One profile
- One slightly low angle and one slightly high angle
- One full-body or mid-shot for proportion and wardrobe
- One candid expression shot (laughing, frowning) for range
This gives the encoder pose variety while keeping identity constant. Ten near-identical frontals do far less work than six well-chosen angles.
Match the lighting and grade
Inconsistent lighting confuses a fusion model more than inconsistent angles. If three references are warm candle-lit and two are cool daylight, the model may treat skin tone as variable and drift toward a mid-tone that matches neither.
Normalise before you upload: similar white balance, similar contrast, similar saturation. If your story needs a night scene, light the night scene in-camera later — do not bake darkness into your identity references.
Remove contradictions
Every disagreement in your reference set becomes a coin flip in the output. Different hairstyles, glasses in some frames but not others, a beard that appears and disappears. Decide the canonical version and exclude the rest.
A useful test: lay all references side by side at thumbnail size. If a stranger could not tell they are the same person, the model will not either.
A repeatable fusion workflow
Here is a workflow that holds up across projects, from a 15-second social clip to a multi-scene narrative piece.
Step 1 — Generate or collect candidates
Start in a controlled space. Use Create Image to build your reference portraits from a consistent prompt seed, or bring in photography if you have it. The advantage of generating them is total control over lighting and wardrobe; the advantage of photography is real skin texture.
Step 2 — Curate ruthlessly
Delete anything blurry, heavily compressed, or stylistically inconsistent. If you would not use it in a portfolio, do not use it in a reference set.
Step 3 — Fuse and lock
Combine the set into a single identity profile. Once locked, treat that profile as immutable for the duration of the project. Changing references mid-project is the single most common cause of drift.
Step 4 — Test in the hardest shot first
Before generating a full scene, generate the shot you are most worried about — usually a profile, a wide, or a strong action pose. If identity holds there, the rest of the sequence will be easy. If it fails, adjust references now rather than after twenty clips.
Step 5 — Generate the sequence in order
Generate shots sequentially and keep the previous shot's final frame as a visual anchor where your tool supports it. Sequential generation gives you a continuity trail you can trace when something goes wrong.
Step 6 — Grade, then review
Apply one consistent grade across all clips at the end. A uniform look hides micro-deviations in skin tone and light that are obvious when clips sit side by side ungraded.
Prompting for continuity across shots
Consistency is not only visual. Prompt structure carries as much weight as your reference images, and sloppy prompts reintroduce the drift you just eliminated.
Describe the person once, reuse the description verbatim
Write a short identity block — age range, build, hair, distinguishing features, wardrobe — and paste it into every shot prompt word for word. Paraphrasing between shots introduces variation. Models are literal; your description should be too.
Keep it under 30 words. Long identity blocks dilute the strongest features.
Change only what the shot requires
A shot prompt should differ from its neighbours in exactly three things: framing, action, and lighting. Everything else stays fixed. This makes drift easy to diagnose — if identity shifts, you know which variable moved.
Use motion language that preserves faces
Aggressive motion is where identity dies. Fast head turns, extreme close-ups during movement, and rapid camera whips give the model fewer frames to maintain structure. Prefer:
- Slow pushes and pulls
- Walking toward or away from camera
- Turns that resolve into a held frame
- Cuts between stable compositions rather than continuous acrobatics
If you need a dynamic shot, generate it slightly wider than you plan to use and reframe in the edit. Cropping in post preserves more facial structure than asking the model to hold a tight frame through movement.
Keyframe control and interpolation
Most drift happens in the gaps between shots, not inside them.
Anchor first and last frames
Where your tool supports it, define the first and last frame of each shot. The model then interpolates rather than inventing. Give it a last frame that matches the first frame of the next shot and you get a continuity bridge that reads as intentional editing rather than an accident.
Match on action, not on position
In classical film grammar, you cut on movement so the eye completes the transition. AI video rewards the same discipline. Cut while the character is turning, or as a hand passes the lens, and small identity deviations get absorbed by the motion.
Know when interpolation breaks down
Interpolation fails predictably:
- When the first and last frames describe different clothing
- When the character passes behind an occluding object and re-emerges
- When lighting changes drastically mid-shot
- When the pose difference is so large there is no believable path between them
Plan these moments as separate shots. It is cheaper to add a cut than to fight a broken interpolation.
Common mistakes and how to avoid them
| Mistake | Symptom | Fix |
|---|---|---|
| Too many similar references | Face locks but feels stiff | Add angle variety, remove duplicates |
| Mixed lighting in references | Skin tone drifts shot to shot | Normalise white balance and contrast |
| Rewriting the identity prompt | Character ages between shots | Copy the identity block verbatim |
| Changing references mid-project | Sudden, unexplained drift | Lock the profile before scene one |
| Pushing for extreme close-ups | Distorted features | Shoot wider, crop in the edit |
| No test shot | Failure discovered late | Generate the hardest shot first |
Most of these are preparation errors, not generation errors. Spending twenty minutes on the reference set saves hours of regeneration.
Quality control across a full sequence
Review the assembled sequence, not individual clips. Play it at normal speed, then at half speed, then scrub through stills.
Ask three questions:
- Identity. Is this recognisably the same person in every shot?
- Silhouette. Does the body shape and wardrobe read consistently in wide shots?
- Continuity of light. Does the grade feel like one film, or several?
When something fails, resist the urge to regenerate the whole scene. Isolate the failing shot, check whether it deviates from your shot-prompt template, and fix that one variable. Regenerating everything resets your continuity trail and usually produces new, different drift.
For teams working across multiple projects, store your locked identity profiles and identity blocks in a shared template so every contributor starts from the same anchor. You can browse reusable starting points in Templates and keep prompt patterns consistent across a series.
Where this fits in a real production
Consider a four-scene short film: a character wakes, walks through a market, argues with a friend, and leaves.
Scene one needs extreme close-ups of a face at rest. Scene two needs full-body movement through a crowd. Scene three needs two characters in conversation with matching eyelines. Scene four needs a wide, backlit exit.
Each of those places a different demand on identity. The close-up exposes facial structure; the crowd scene exposes wardrobe and proportion; the conversation exposes that both characters must hold steady simultaneously; the backlit exit exposes silhouette.
If you build your reference set to cover all four demands before generating shot one, the sequence holds together. If you build it from portraits alone, scene two and scene four will break, and you will be tempted to blame the model rather than the inputs.
The same logic scales down. A 20-second product spot with a recurring presenter, an explainer with a consistent avatar, a serialised social format with the same host — all of them benefit from a locked identity profile. If you are producing at volume, start from Create Video and treat the identity profile as part of your project setup, not an afterthought.
FAQ
How many reference images do I actually need?
Six to twelve well-chosen images covering front, both three-quarter angles, profile, and one wider shot. More than twelve rarely helps and often hurts by diluting the signal with near-duplicates.
Can I use one reference image and just write a better prompt?
You can, and for tight, simple shots it sometimes works. But a single image gives the model exactly one pose and one lighting condition to learn from, so any shot that departs from it invites invention. Multiple references are the more reliable path.
Why does the character change after a cut even though nothing changed in my prompt?
Usually a seed, a reference weighting, or a subtle lighting description changed. Check your shot-prompt template line by line. The second most common cause is a scene-level lighting instruction that overrides your identity block.
Should I include different expressions in the reference set?
One or two, at most. Expression is part of what gets encoded, so a set full of laughter can produce a character who smiles in dramatic scenes. Keep most references neutral and control expression through the shot prompt.
How do I handle a character who changes costume during the story?
Lock the identity from face-focused references and treat wardrobe as a separate layer described in the shot prompt. Keep at least one full-body reference per costume so proportion stays consistent.
What about multiple characters in one shot?
Fuse each character separately, then describe them distinctly in the prompt — position, action, and one memorable feature each. Avoid two characters with similar hair colour or build; ambiguity is where fusion collapses.
Does image fusion work for stylised and animated looks?
Yes, and often more easily, because the model has fewer photoreal micro-details to get wrong. The principles are identical: cover angles, normalise the look, lock the profile, test the hardest shot first.
Start with your reference set, not your prompt
The temptation with AI video is to treat generation as the creative act and preparation as overhead. In practice, preparation is the creative act. A locked identity profile, a consistent identity block, and a shot list ordered from hardest to easiest will do more for perceived production value than any amount of prompt tinkering.
Build the reference set. Fuse it once. Reuse it everywhere. Then open Orelon and put your consistent character into motion — because the difference between a demo and a film is not the model, it is whether the same person walks out the other side of the cut.