Build Consistent LEGO-Pixel Characters with Multi-Image AI Video

Why AI Video Loses Character Identity Between Shots

Text-to-video models are extraordinary at inventing a moment. They are much weaker at remembering one. A prompt like "a blocky pixel knight in red armor walks through a market" describes a category of person, not an individual. Shot one renders one knight. Shot two, generated minutes later with a slightly different camera angle and lighting description, renders a plausible cousin: the armor is a shade more orange, the helmet crest is gone, the shoulders are wider, and the face plate has shifted from a visor to a grille.

None of these changes matter in isolation. Together they destroy the illusion of a continuous story. Viewers forgive soft lighting and imperfect motion. They do not forgive a protagonist who becomes a different person every four seconds.

This is fundamentally a memory problem, not a rendering problem. Video diffusion samples noise conditioned on your text and a few reference signals. Small latent differences compound frame by frame and shot by shot. Unless you deliberately constrain the identity signal, the model will happily re-interpret your character on every generation. The fix is not a better adjective. The fix is a reference system: multiple images of the same character fed into the generation as an identity anchor, plus a workflow that keeps those anchors stable across an entire sequence.

Multi-image fusion is the practical name for that system. Instead of conditioning a shot on a single still, you supply a small, curated set of views of the same character — frontal, three-quarter, profile, back, maybe a close-up of the helmet — and the model blends those signals into a consistent internal identity. The rest of this guide is about how to build that set, how to weight it, and how to keep it working when your project grows from one test clip into a full series.

The LEGO-Pixel Aesthetic: Helpful, and Treacherous

Blocky figure aesthetics are one of the friendliest styles for generative video, for three reasons.

First, stylization raises viewer tolerance. Nobody expects a minifigure-like character to have photoreal skin pores, so small deviations in geometry read as style rather than error. Second, the geometry is modular: cylinders, studs, C-clip hands, flat plates. Modular forms are easier for a model to reproduce than organic anatomy. Third, the palette is usually limited and graphic, which means color can carry a lot of identity weight.

But there is a trap. The same qualities that make the style forgiving also make characters harder to tell apart. Two blocky knights in similar armor look like the same asset with different lighting. When every face is a flat printed smile, you lose the single most reliable identity anchor in realistic video: facial structure.

So in a LEGO-pixel project, identity has to be concentrated elsewhere. Build your character around at least four strong, redundant cues:

Silhouette: helmet shape, crest, horns, hat brim, cape outline, backpack bulk.
Color blocking: a specific torso print, a two-tone arm split, an unusual leg color, a scarf in a saturated accent hue.
Signature prop: a wrench, a lantern, a shield with a fixed emblem, a bandolier.
Micro-detail: a chipped corner on the helmet, a mismatched hand color, a sticker on the torso.

Write these cues down as a locked list. Every prompt for that character must include at least three of them, in the same order, using the same words.

Decoding the Visual Grammar Before You Generate

Before you build a reference pack, spend ten minutes writing your style rules. Ambiguity here is the root cause of most drift later.

Pixel scale and geometry

Decide how literal the pixel treatment is. Options range from "clean 3D blocky render with no pixelation" to "heavy dithering on a low-resolution grid." Pick a grid size and commit. Mixing a 64-pixel-wide character with a 256-pixel-wide environment looks accidentally broken rather than stylized.

Proportion contract

Blocky figures have fixed ratios. Head width relative to torso, arm length relative to leg length, hand size relative to forearm. If your model drifts during a walking sequence — suddenly the arms get longer — the character stops being recognizable even if the colors are perfect. State the proportion rules in your style bible and repeat the key ones in prompts.

Lighting rig

Pick one lighting setup for the whole project: a key light from the upper left, soft fill from the right, no rim light. Multi-image fusion works dramatically better when all references share a lighting setup, because the model does not have to disentangle shadows from geometry. Lighting variation is a post-production decision, not a reference-pack decision.

Palette lock

Define your character's colors as hex values in a text file. Then translate them into words you will reuse identically: "brick red," not "red" in one prompt and "crimson" in the next. Verbal synonyms are a hidden source of drift because they shift the sampled color distribution.

Building a Character Reference Sheet That Actually Works

Your reference pack is the memory of your project. Treat it like a casting document, not a mood board.

A strong pack contains five to eight images:

Neutral front view, full body, flat grey background.
Three-quarter view, same scale, same lighting.
Profile view.
Back view, especially important for capes and backpacks.
Close-up of the head and helmet.
One action pose that shows how joints behave at speed.
Optional: a second expression or visor state.

Rules that matter more than the count:

Consistent headroom and framing. Crop every reference so the character occupies the same percentage of the frame.
One background. Flat mid-grey. Backgrounds leak color into characters during fusion.
One lighting rig. As above.
No occlusion. Do not let a hand cover the torso print in the only view that shows it.
Name everything. hero_front_v3.png, not final_final2.png.

More references are not better. Beyond roughly eight images, you start feeding the model contradictory or redundant signals, and fusion begins averaging toward a generic figure. Curate ruthlessly: if a reference does not add a distinct angle or a distinct detail, delete it.

How Multi-Image Fusion Works in Practice

The mechanical idea is straightforward. A fusion-capable video model attaches an identity embedding to each reference image, then blends those embeddings while conditioning a shot. The blend is not an average of pixels — it is an average of features — and that distinction explains most of the technique's behavior.

Anchors versus accents

Divide your references into two roles. Anchors define who the character is: the front view and the three-quarter view, weighted highest. Accents define details: the helmet close-up, the back view, the prop shot. Accents should inform the generation without dominating the silhouette.

If your tool lets you weight references, start with anchors at high weight and accents at roughly half. If your tool blends equally, control emphasis through the prompt: describe anchor-derived attributes first in the sentence, since early tokens typically carry more conditioning influence.

Identity survives in features, not in pixels

Because fusion operates on features, it is robust to changes of pose, camera, and background — and fragile to changes of style. Feeding a photoreal reference image into a project that renders blocky pixel characters produces a hybrid with realistic shading on blocky geometry. That hybrid is usually worse than either pure style. Keep every reference in the target style.

Shot generation strategies

There are three common ways to apply fusion across a sequence:

Fresh generation with the full reference pack. Maximum flexibility, maximum drift risk. Use it for establishing shots and environments.
Image-to-video from a locked anchor frame. Take one approved still of the character in the target pose and animate from it. This is the highest-consistency method for dialogue and close-ups.
Video-to-video style pass. Generate motion in a simpler render, then restyle the whole clip in one pass so the character is treated uniformly. Excellent for long action sequences.

Most reliable productions mix all three, choosing per shot type rather than per project.

A Step-by-Step Workflow from Script to Locked Character

Step 1: Write the cast bible

One page per character. Cues, palette hex values, proportion rules, signature prop, forbidden variations ("never give him a cape"). This document is the source of truth that keeps prompts consistent weeks later.

Step 2: Create the hero still

Generate many candidates for a single, neutral, well-lit still of each character. Iterate on that one image until you would be happy to see it as a poster. Everything downstream inherits its flaws, so this is the step worth over-investing in.

Step 3: Expand into a reference pack

Use the hero still as a reference and generate the turnaround views, then the helmet close-up. Fix small inconsistencies by hand-editing rather than regenerating from scratch — a corrected profile view is more valuable than a fresh lucky roll.

Step 4: Generate shot by shot, anchor first

For each scene, generate the widest, most informative shot first. Approve it. Then use it as a reference frame for the closer shots. Working wide-to-tight keeps the spatial logic and the lighting consistent, and it means every tight shot inherits an approved identity.

Step 5: Apply a uniform style pass

If any shot came out with slightly different pixel density or shading, run the whole sequence through one style pass at the end. Uniform imperfection looks intentional; mixed perfection looks broken.

Step 6: Review with a checklist, not a feeling

Do not ask "does this feel right?" Ask specific questions: Is the helmet crest present? Is the torso print the same orientation? Are the hands the correct color? Did the cape length change between shots? A checklist catches drift that a tired eye misses.

Prompt Patterns That Preserve Identity

Structured prompts outperform poetic ones for consistency work. Use six blocks, always in the same order:

Style block: identical wording in every prompt for the project.
Identity block: the character's locked cues, in a fixed order.
Action block: what the character is doing.
Camera block: framing, angle, focal length language.
Lighting block: matching your locked rig.
Negative block: the failure modes you have actually observed.

A concrete example of an identity block: "blocky pixel minifigure-style knight, brick red torso print with silver chevron, silver visored helmet with a single left crest, mismatched dark grey left hand, short black cape, squat proportions with wide shoulders."

Reuse that exact string — not a paraphrase — in every prompt for that character. It sounds mechanical. It is mechanical. That is the point.

Keep camera language equally frozen. If you describe the lens as "wide angle, low angle" in shot one, do not switch to "telephoto, eye level" unless the story requires it, because lens changes alter apparent proportions and your model will happily reshape the character to match.

Negatives worth defaulting to: "no realistic skin texture, no extra fingers or studs, no glowing outlines, no background characters touching the subject, no color bleeding from background."

Choosing Tools for Consistency-First Productions

Not every generator is built for this work. Evaluate candidates against these criteria rather than against raw visual wow factor.

Reference count and weighting. Can you supply five or more images and control their relative influence?
Identity adapters. Does the tool support a reusable identity object, so you do not re-upload references every run?
Seed control. Can you lock a seed and reproduce a shot almost exactly with one changed parameter?
Motion quality at short durations. Consistency collapses when motion smears. Short, controlled clips stitch better than long ones.
Style transfer pass. Is there a whole-sequence restyle option?
API and batch access. A series needs programmatic runs, not one-by-one clicking.
Output determinism. Run the same job twice. How different are the results? Low variance is worth more than peak quality here.
Licensing clarity. Especially if characters will appear in monetized work.

A useful test: pick your hardest character — the one with the most detail — and generate the same five-shot mini-scene in three candidate tools. Score identity retention, palette stability, and motion. The winner is often not the tool with the prettiest demo reel.

Common Mistakes and How to Fix Them

Too many references. Symptom: the character looks slightly generic and no single detail is crisp. Fix: cut to five images, keep two anchors and three accents.

Conflicting references. Symptom: the helmet flickers between two shapes across a clip. Fix: version your pack, and delete superseded images instead of leaving them in the folder.

Prompt synonym drift. Symptom: armor color shifts across shots. Fix: one canonical adjective per color, copied and pasted.

Changing style mid-project. Symptom: some shots look pixelated, others smooth. Fix: freeze the style block before shot one and run a final uniform style pass.

Fixing drift with upscaling. Symptom: a wrong-looking character rendered beautifully at high resolution. Fix: upscale last, never first. Resolution does not repair identity.

Ignoring the environment. Symptom: a warm sunset background tints your red armor orange. Fix: restrict environmental color language, and keep backgrounds flat in references.

Generating the hero shot last. Symptom: you discover in editing that the best-looking character was created for the final scene and matches nothing earlier. Fix: always lock the hero first.

Scaling from One Clip to a Series

A single consistent character is a technical achievement. Twelve consistent characters across forty shots is a production system. Four practices make the difference.

Asset discipline. One folder per character, containing the cast bible, the locked reference pack, approved hero still, and an output log recording which seed and prompt produced which approved shot. When you need shot 37 to match shot 12, that log is the only thing that saves you.

Template prompts. Store prompt skeletons with empty slots for action and camera. This removes the temptation to improvise wording under deadline pressure.

Review gates. Approve at three points: hero still, reference pack, and wide shot of each scene. Catching drift at the wide-shot stage costs one regeneration. Catching it after the close-ups costs a day.

Reusable environments. Lock background plates, props, and crowd assets the same way you lock characters. Environment color drift is one of the most common causes of character color drift, because the model rebalances the whole frame.

FAQ

How many reference images does a character actually need?

Five to eight for most projects. Two strong anchors, three to five accent views. Beyond eight, fusion tends to average toward a generic figure rather than sharpening identity.

Can I get consistent characters from text prompts alone?

For very short clips, sometimes. Across a sequence, no. Text describes categories. Identity requires visual references that survive shot boundaries.

Why does my blocky character keep changing color?

Almost always one of three causes: inconsistent adjectives across prompts, a colorful background bleeding into the character, or a reference pack with mixed lighting. Fix the adjective list first, then the references.

Is a blocky pixel style easier or harder than photorealism?

Easier to look acceptable, harder to make distinctive. Geometry is forgiving, but identity has to be carried by silhouette, color, and props instead of facial features.

Should I generate long clips or short ones?

Short clips, then stitch. Consistency degrades with duration because every additional second is another opportunity for the identity signal to drift.

How do I keep two similar characters apart?

Give them different silhouettes before you give them different colors. If both wear the same helmet shape, viewers will confuse them even if one is red and one is blue. Then differentiate palettes, then props.

What if the character looks right but the motion looks wrong?

Separate the problems. Generate an approved still, animate from it with restrained motion, and restyle the sequence in one pass at the end. Trying to solve motion and identity in a single generation usually solves neither.

A Final Checklist Before You Render

Cast bible written, with locked cues and hex colors.
Hero still approved for every character.
Reference pack at five to eight images, one lighting rig, one background.
Style block frozen and copied identically into every prompt.
Identity block uses the same words in the same order.
Widest shot of each scene generated and approved first.
Consistent shot types mapped to consistent generation strategies.
Review gate checklist in hand, not a vague sense of unease.
Uniform style pass scheduled for the finished sequence.
Output log updated so the next session can reproduce today's work.

Consistency in generative video is not a single clever prompt. It is a small, disciplined library of visual facts that you refuse to improvise with. Build the library once, protect it obsessively, and the blocky pixel heroes of your series will stay recognizably themselves from the opening frame to the last.