Consistent characters are the difference between a demo clip and an actual story. As soon as an AI-generated video moves from one shot to the next, every shortcut in your reference setup becomes visible: a jawline softens, a coat shifts shade, a scar migrates to the wrong cheek. Multi-image fusion — conditioning a video model on several reference images of the same subject at once — is the most dependable way to keep that identity anchored across scenes, angles, and lighting changes. This guide explains how the technique works, how to build reference packs that survive real production, and the workflow that turns consistency into a repeatable habit rather than a lucky streak.
The Real Problem: Identity Drift Across Scenes
Text-to-video models do not remember your character. They sample. Every generation is a fresh draw from a probability distribution shaped by your prompt, the model's training data, and the random seed. A prompt like "a woman in her thirties with short dark hair and a red coat" describes an archetype, not a person. Run it three times and you get three different women who all match the description.
Drift compounds when the scene changes. A new location means new light, a new lens, a new pose, and often a new emotional register. Each of those variables pushes the model to re-interpret the face, and small deviations accumulate. Scene one gives you a 32-year-old with a narrow chin. By scene four the same character reads as 45, with a wider jaw and lighter hair. If someone is watching a single clip, they may not notice. Across a sequence, the illusion collapses instantly.
There is a second, subtler form of drift: wardrobe and props. A jacket that starts as burgundy becomes maroon, then plain red. A pendant moves from left to right. A coffee cup changes shape between cuts. Audiences forgive a lot, but they are extremely good at noticing objects and accessories that refuse to stay put, because those details are how we track continuity in real films.
The practical takeaway is that consistency is not a property you can prompt into existence with adjectives. It has to be supplied as data — as images the model can fuse together.
How Multi-Image Fusion Actually Works
Multi-image fusion means giving the model more than one visual anchor for the same subject. Instead of a single portrait, you provide a small set: front, three-quarter, profile, a full-body shot, a detail of the face, maybe an expression variation. The model encodes each reference into an identity representation and blends them during generation, so the output is pulled toward a shared, averaged identity rather than any single frame.
Identity pack versus single reference
A single reference image is the weakest possible setup. It locks one angle, one lighting setup, and one expression. The moment your scene asks for a different camera angle, the model has to invent the parts it cannot see, and invention is where drift starts. A single reference also tends to over-copy incidental features — the exact background, the exact shadow under the nose — which creates a different problem: the same portrait appears in every scene with a different backdrop pasted behind it.
An identity pack of four to eight images solves both issues. It gives the model enough viewpoints to infer the underlying structure of the face, and enough lighting variety to separate "who this person is" from "how this person was lit."
The three layers of a fused identity
Think of a character as three stacked layers, each with different stability requirements.
Face and skull structure is the most sensitive layer. Eye spacing, nose bridge width, jaw shape, brow height, and skin texture must match closely. Small errors here read as a different person.
Hair, wardrobe, and accessories is a mid-sensitivity layer. You can tolerate minor fabric folds and lighting shifts, but color, silhouette, and key details like a collar shape or a specific watch must stay stable.
Body proportions and posture is the loosest layer. Height, shoulder width, gait, and habitual gestures can vary slightly between shots without breaking continuity — and they often should vary, because real performers move.
When you build references, prioritize the top layer. Two excellent face references beat ten mediocre full-body shots.
Keyframes and motion control
Fusion handles identity; keyframes and motion controls handle continuity of action. A keyframe gives the model a fixed starting or ending pose, which is especially useful for shot-reverse-shot dialogue, match cuts, and any sequence where a character has to be in a specific place at a specific moment. Motion control — whether through image-to-video with a start frame, a pose guide, or a camera-path specification — keeps the body moving the way you want while the fused identity keeps the face stable.
The workflow implication: decide which problem you are solving before you generate. If the face keeps changing, add references. If the face is stable but the movement is wrong, add keyframes and camera direction.
Building an Identity Pack That Survives Every Scene
Reference coverage: what to include
A reliable pack for a recurring character usually contains:
- One clean, evenly lit front-facing portrait, eyes open, neutral expression
- One three-quarter turn, roughly 30 to 45 degrees off axis
- One profile or near-profile view
- One shot in harsher or directional light to teach the model how the face reads in contrast
- One full-body or knee-up shot for proportions and default wardrobe
- One expression variation, such as a laugh or a frown, if the character has emotionally varied scenes
- Two to four close crops of the eyes, nose, and mouth if the character has a distinctive feature
Reference hygiene
Reference quality matters more than reference count. Blurry, over-compressed, or heavily filtered images inject noise into the identity embedding. Avoid watermarks, busy backgrounds, and heavy skin smoothing — a model that learns "poreless plastic" will render poreless plastic in every scene. Keep lighting consistent enough to describe one person, but varied enough to separate identity from illumination. Crop tightly; a reference where the character occupies a fifth of the frame wastes most of its detail budget.
Also check for conflicting signals. If one reference shows a beard and another is clean-shaven, the model will blend them and produce stubble in every shot. If one reference has the character smiling and another scowling, expressiveness may bleed into scenes where it does not belong.
The identity contract
Write a short, reusable text block and paste it into every prompt. It should describe only what images cannot reliably convey — personality, demeanor, and the details you want held steady in words. Keep it under 60 words and never contradict the references.
Example block:
ANA — 5'7", lean build, close-cropped dark hair, deep-set brown eyes, a thin scar through the left eyebrow, olive skin. Quiet, watchful, rarely smiles. Wears an oversized charcoal wool coat over a plain white tee, silver ring on the right index finger. Neutral camera framing unless specified.
Reuse this exact block verbatim across scenes. Paraphrasing between shots reintroduces the drift you are trying to eliminate.
A Step-by-Step Workflow for a Five-Scene Sequence
Step 1: Write the character bible
Before generating anything, write down the facts: name, age range, ethnicity, build, hair, eyes, defining features, wardrobe, and props. Decide which details are load-bearing — a scar, a specific jacket, a wedding ring — and which are negotiable. This document becomes your quality standard later, so be specific.
Step 2: Generate and approve a hero reference
Generate a single high-quality still of the character in neutral conditions. Treat this as a casting decision, not a draft. Iterate until you would be happy to see this face for an entire film. Everything downstream inherits its strengths and flaws.
Step 3: Expand into a multi-angle pack
Use image-to-image, an image editor, or a character-sheet generator to produce the other angles from the hero reference. Keep the same hair, same wardrobe, same lighting family. Inspect each new image against the hero: if the jaw or eye spacing changes, discard it. A slightly boring but accurate pack outperforms a visually exciting inconsistent one.
Step 4: Define visual grammar per scene
For each scene, write down the location, time of day, lighting direction, lens character, camera movement, and the emotional beat. This is where you decide how much variation is legitimate. A night interior can legitimately darken skin tones; it cannot legitimately change bone structure.
Step 5: Generate shot by shot, fusing the pack
Feed the identity pack plus the scene prompt plus any keyframes. Generate short clips rather than full scenes — three to six seconds each — so a bad take stays cheap to discard. Approve each shot before moving on; fixing drift early is far easier than fixing it during the edit.
Step 6: Assemble, then repair surgically
Cut the approved shots together. When you spot a mismatch, do not regenerate the whole scene. Identify whether the problem is identity, wardrobe, lighting, or motion, then fix only that layer — a single regenerated shot with a tightened reference set is usually enough.
Prompt Patterns That Improve Fusion Accuracy
Structure your prompt in a fixed order so the model receives consistent signal every time: subject block, action, camera, lighting, environment, style. Keep style tokens identical across the sequence. If scene one says "cinematic, 35mm, shallow depth of field," scene four should say the same thing unless the change is intentional.
Use explicit identity anchors sparingly but consistently: hair length, hair color, one distinguishing feature, and one wardrobe item. Do not list ten facial attributes; that produces a generic composite. Let the images carry the detail.
Avoid negative framing where possible. "Not blonde" is weaker than "dark brown hair." Prefer positive descriptions, and reserve negatives for artifacts like extra fingers or text overlays.
Finally, control the seed when your tool allows it. Reusing a seed across shots in the same scene reduces the random variation the model introduces on its own.
Handling Props, Vehicles, and Recurring Objects
Characters are not the only things that need continuity. A recurring object — a motorcycle, a specific book, a piece of jewelry — benefits from the same treatment. Build a small object pack with two or three angles, then reference it in every shot where it appears.
For objects that are manipulated on camera, add a keyframe for the hand position. Hands are the most common failure point in AI video, and a prop gives the model something concrete to hold, which paradoxically improves hand anatomy.
Name props explicitly in the prompt and keep the wording identical. "A scuffed silver lighter" stays consistent; "a lighter," "her lighter," and "the metal lighter" invite three different objects.
Common Failure Modes and How to Fix Them
| Symptom | Likely cause | Fix |
|---|---|---|
| Face changes between shots | Too few references or conflicting references | Add angles, remove outliers, tighten the identity block |
| Same portrait in every scene | Single reference dominating | Add lighting variety, reduce reference weight if the tool exposes it |
| Wardrobe shifts color | Vague color language or mixed lighting | Specify precise color words, keep style tokens fixed |
| Age drifts older | Harsh shadows and high-contrast grading | Soften key light, regenerate flatter, then grade in post |
| Hair length changes | Inconsistent references | Remove cropped or partially occluded hair references |
| Body proportions shift | No full-body reference | Add one knee-up or full-body shot to the pack |
The pattern behind most failures is the same: the model is averaging conflicting inputs. Fix the inputs before you fix the prompt.
Tooling Landscape: Choosing the Right Pipeline
Most modern video generators accept multiple reference images in some form, but the implementation varies. Some expose a dedicated character or element slot; others expect you to attach references inside the prompt; a few rely on separate identity-adaptation steps you run before generation.
Choose based on three questions. First, how many references can the tool fuse at once, and does it weight them? Second, does it support keyframes or start-frame conditioning for motion control? Third, how long are the clips, since longer generations tend to drift further from the reference.
Practical advice: keep a generative image tool for building the identity pack, a video generator for motion, and a compositor such as DaVinci Resolve or After Effects for assembly and color matching. Trying to do everything in one tool usually means compromising on the pack quality that consistency depends on.
Quality Checklist Before You Publish
Run through this before exporting:
- Watch the sequence at normal speed, not frame by frame. Drift that is invisible in stills can be obvious in motion, and vice versa.
- Freeze on every cut and compare the character's face, hair, and wardrobe to the bible.
- Check screen direction and eyelines — continuity errors there feel like identity errors to viewers.
- Verify lighting continuity across cuts within a single scene.
- Confirm props appear in the same hand, same size, same color.
- Mute the audio and watch once. Visual continuity problems are much easier to catch without dialogue.
FAQ
How many reference images do I actually need?
Four to eight is the sweet spot for most models. Fewer than four usually means you are asking the model to invent too much; more than ten rarely improves results and can dilute the identity signal if any reference is off-model.
Can I fix consistency in post instead of regenerating?
Sometimes. Color matching, minor retouching, and frame blending can hide small deviations. Structural problems — a changed jawline or a different nose — cannot be fixed convincingly in post, so regenerate those shots.
Does multi-image fusion work for stylized or animated characters?
Yes, and it often works better, because stylized characters have fewer ambiguous features. The main adjustment is that your references should share the same art style, since the model fuses stylistic choices along with identity.
Why does my character look better in wide shots than close-ups?
Close-ups expose every inconsistency in the identity embedding. If your close-ups drift, add high-resolution face references and prefer slightly wider framing until the pack stabilizes.
Should every scene use the full reference pack?
Yes for scenes featuring the character's face. If a scene is a distant silhouette or a shot from behind, you can relax the pack, but keeping it attached costs nothing and protects against surprises.
How do I keep a character consistent across different projects or episodes?
Freeze the pack. Store the reference images, the identity text block, and the style tokens together as a versioned asset. When you revisit the character, reuse the exact same files rather than generating similar ones from scratch. That single habit eliminates most long-term continuity problems before they start.
