Ask anyone who has spent a week building a narrative with generative video what the hardest part is, and you will rarely hear "render quality." The real bottleneck is continuity: keeping a character recognizable from the first frame to the last. A face drifts. A jacket changes from olive to brown. Hair length quietly shifts between cuts. The viewer may not name the problem, but they feel it instantly, and the story stops being about a person and becomes about a model's mood swings.
The fix is not a single magic prompt. It is a production discipline built on reference images: collecting several well-chosen images of the same character and merging what they teach the model into one stable identity that survives dozens of shots, camera moves, and lighting changes.
This guide walks through that discipline end to end. You will see how to assemble a reference pack, how to merge multiple images into one usable character profile, how to prompt and control generation so the face holds, and how to choose between tools that handle this well. It is written for short films, episodic series, brand content, explainers, and game cinematics where the same person has to stay the same person.
Why character consistency breaks in generative video
A video model does not store a character. It stores a statistical relationship between the pixels in front of it and the words in your prompt. Every time you generate a new shot, the model re-invents the face from scratch, guided only by whatever conditioning you supply. If your conditioning is a sentence like "a woman in her thirties with dark curly hair," the model is free to interpret that sentence differently in every clip. Across ten shots, you get ten cousins.
Three forces make this worse as a project grows.
Weak identity signal. Text prompts describe categories, not people. "Sharp jawline" and "warm smile" are adjectives, not biometrics. Without images, the model has no anchor for the specific arrangement of features that makes your protagonist look like your protagonist.
Temporal drift inside a clip. Even within a single generation, models trade identity for motion. When a character turns their head quickly or passes through a shadow, the model may re-synthesize facial structure mid-shot. The result is a subtle morph that viewers read as uncanny even when they cannot explain why.
Cross-shot variation. Different camera angles, focal lengths, and lighting setups demand different information from the model. A profile view needs jaw and ear geometry. A low-angle shot distorts proportions. If your reference material only covers a straight-on portrait, the model improvises everything else, and improvisation produces inconsistency.
Understanding these three failure modes matters because they require different remedies. The first is solved by reference images. The second is solved by motion control and shorter shots. The third is solved by a reference pack that covers the angles and lighting conditions your story actually uses.
What merging multiple images really means
"Merging images" sounds like a post-production operation in an image editor. In practice it is a conditioning operation: you are combining information from several stills so that a single latent identity emerges, richer than any one photo could produce.
Think of each reference image as answering one question the model will otherwise guess at. A neutral front-facing portrait answers "what does this face look like at rest?" A three-quarter view answers "how do the cheekbones and nose read from an angle?" A full-body shot answers "what is the body proportion and silhouette?" A candid with an unusual expression answers "what happens to the mouth and eyes under emotion?" A low-light frame answers "what does this skin tone do when the key light is warm and off-axis?"
When you supply one image, the model overfits to it. You get excellent likeness in frames that resemble that photo and strangers everywhere else. When you supply five to eight complementary images, the model can triangulate. The identity becomes a region of possibility rather than a single point, and that region is stable enough to survive a new camera angle.
There is a ceiling. Too many images, especially contradictory ones, blur the identity rather than sharpen it. If your pack contains two different hairstyles, three distinct clothing sets, and inconsistent lighting, the model averages them and produces a face that resembles nobody. The craft is in curation, not volume.
Building a reference pack that actually works
A good reference pack is assembled with the same rigor as a casting session. Aim for six to ten images, all treated as canon.
Cover the angles you will shoot
Start with a clear, evenly lit front view. Add a three-quarter left and three-quarter right. Add one profile if your script includes profile shots. Add one mild high angle and one mild low angle if your shot list uses them. Angles you never shoot do not need coverage, but angles you shoot without reference will drift.
Lock wardrobe and hair per sequence
Decide early whether your character has one canonical look or several. If a story spans seasons or timelines, create a separate pack per look and label them clearly. Mixing looks inside one pack is the single most common cause of a mushy, unrecognizable face.
Keep expression varied but plausible
Include a relaxed neutral, a genuine smile, and one focused or serious expression. Avoid extreme grimaces and heavy stylization, since those frames pull the identity toward caricature.
Normalize technical quality
Crop tightly enough that the face occupies a meaningful portion of the frame. Avoid heavy filters, beauty retouching, motion blur, and strong color grading. Keep resolution consistent. A pack that mixes a crisp studio portrait with a compressed screenshot will teach the model to reproduce compression artifacts as skin texture.
Write down what each image teaches
A short reference sheet — one line per image describing angle, expression, wardrobe, and lighting — saves hours of debugging later. When a shot drifts, you can trace which reference was underweighted and rebalance.
The identity merge workflow, step by step
Once the pack exists, the merging happens across several passes rather than in one action.
Step 1: Generate an identity board
Use an image model to synthesize a single canonical portrait that combines the best traits from your references — the jawline from one, the eye shape from another, the hair texture from a third. Iterate until you have a portrait that looks like a plausible person and looks like all of your references simultaneously. This portrait becomes your anchor.
Step 2: Expand the anchor into a turnaround
From the anchor portrait, generate a set of consistent views: front, three-quarter, profile, and full body. Fix any drift by hand in an image editor rather than regenerating endlessly. Each corrected view strengthens the identity you will carry into video.
Step 3: Test the identity against hard conditions
Before committing to a full shot list, run short video tests: a slow head turn, a walk toward camera, a shot with strong backlight, a shot with a handheld feel. These stress tests reveal whether the identity holds under motion and lighting change. If it fails, go back to the board, not to the prompt.
Step 4: Generate shots with identity conditioning
Feed the canonical images into your video model as image or character conditioning, alongside your prompt. Keep the reference set identical across every shot in a sequence. Changing references mid-sequence is a guaranteed continuity break.
Step 5: Assemble and repair
Cut the shots together early, before polishing. Continuity problems are easiest to see in motion. Repair by regenerating only the failing shot with a tighter reference set, or by compositing a corrected face into the existing take.
Treat this as a pipeline with a single source of truth. Every downstream asset should trace back to the identity board.
Prompting and control techniques that hold a face together
Prompts do not create identity, but they protect it. Several habits make a measurable difference.
Describe what stays constant, not what changes. Keep a fixed block of text describing the character — bone structure, eye color, hair, skin tone, signature wardrobe — and reuse it verbatim in every prompt. Then add only the shot-specific details: camera angle, action, lighting, mood. Varying the character block reintroduces randomness.
Front-load identity tokens. Place the character description in the first sentence. Many models weight early tokens more heavily, and late identity details get diluted by scene description.
Use negative constraints sparingly and specifically. Blanket negatives like "no distortion" rarely help. Targeted ones — "no beard," "no glasses," "no hat" — prevent specific confusions when a prior reference might imply those features.
Prefer shorter shots. Identity decay accelerates with duration. Three four-second shots hold a face far better than one twelve-second shot, and you can join them in the edit.
Control motion deliberately. Fast rotations and large occlusions force the model to hallucinate. Slow the action, keep the face visible, and let the camera do more of the work.
Match lighting to your references. If every reference is soft daylight, a hard neon night shot will push the model into unfamiliar territory. Either add a night reference or expect to repair.
Use a seed and reuse it. Many models let you reuse a generation seed. Reusing it alongside the same references reduces incidental variation between takes.
Scene-to-scene continuity beyond the face
Identity is more than a headshot. Audiences track wardrobe, props, hair, and lighting. Each of these needs the same reference discipline.
Wardrobe should be documented with at least one full-body reference per outfit, plus a note on layering and accessories. Small details — a watch, a scar, a belt — are memorable to viewers and easy for models to drop.
Hands and props deserve their own reference stills. If your character carries an object through a sequence, generate a clean reference of the object held, and reuse it. Props morph as readily as faces.
Lighting continuity is often handled better in post than in generation. Lock a look with a LUT or grade and apply it across the sequence so that minor lighting differences between generated shots disappear.
Geography and blocking matter for longer sequences. Keep a simple floor plan and screen direction notes. If your character exits frame left, they should enter frame right in the next shot unless you intend a deliberate break.
Color of the environment influences perceived skin tone. A character who reads warm and healthy in a sunlit scene can look gray in a cold blue interior. Compensating in the grade is faster than regenerating.
Choosing the right tool for each part of the job
No single tool is best at everything, and the strongest pipelines mix several.
Portrait and identity boards: Midjourney, Flux-based models, and Stable Diffusion with a character or style adapter are all strong. The distinguishing factor is how well the tool responds to multi-reference conditioning. Tools that accept several input images at once tend to produce better merged identities than tools that accept only one.
Image-to-video animation: Runway, Kling, Luma Dream Machine, Pika, and Sora-class models each handle motion differently. Some are excellent at subtle facial performance, others at camera movement. Test the same reference image across two or three and pick per shot type rather than globally.
Character conditioning features: Several platforms now offer a dedicated character or subject reference slot. These are the most reliable option when available, because they were trained specifically to preserve identity rather than to reproduce a single frame.
Iterative node-based workflows: ComfyUI and similar graph tools give you precise control over conditioning strength and reference weighting. They demand more setup but repay it when a project has forty shots that all must match.
Repair and finishing: Photoshop, Krita, or Affinity Photo for stills; DaVinci Resolve, After Effects, or Nuke for compositing and grade; Topaz Video AI or similar for upscaling and deflicker. Post-production is not a failure of generation — it is part of the pipeline.
Practical selection criteria: how many reference images the tool accepts, whether it supports identity conditioning separate from style, how stable it is across seeds, output resolution, and how quickly you can iterate. Test each candidate on the same three-shot stress test before committing a project to it.
Common mistakes and how to fix them
Using too few references. One image produces a rigid identity that only works at one angle. Fix: add three to five complementary angles.
Using contradictory references. Two haircuts or two body types in one pack average into a stranger. Fix: split into separate packs per look and label them.
Changing references mid-sequence. Even small swaps shift the face. Fix: freeze the reference set for the entire sequence and change it only at intentional story breaks.
Overwriting the character block in prompts. Paraphrasing your identity description each time reintroduces randomness. Fix: keep a copy-paste block and only edit the shot-specific tail.
Chasing long takes. Longer generations drift more. Fix: generate shorter beats and assemble them in the edit.
Ignoring the first frame. A weak opening frame poisons everything downstream. Fix: lock a strong hero still before animating.
Fixing everything by regenerating. Endless regeneration burns time and rarely converges. Fix: composite or paint a corrected frame and re-animate from it.
Quality control checklist for every sequence
Before you call a sequence finished, run these checks in order.
Watch the sequence muted and at normal speed. Continuity errors are more visible without dialogue.
Freeze on each cut and compare the character against the identity board. Check eye spacing, nose shape, jawline, hairline, and skin tone.
Verify wardrobe, accessories, and props across every cut.
Check that screen direction and eyelines are consistent.
Look for flicker, warping, or texture crawl on the face, especially during turns.
Confirm lighting direction matches between adjacent shots.
Inspect at full resolution on a calibrated display, not on a laptop at half brightness.
Export a short reference reel of the character's best shots. It becomes the identity pack for the next episode and shortens setup dramatically.
FAQ
How many reference images do I need?
Six to ten well-chosen images covering multiple angles, one or two expressions, and your canonical wardrobe is a solid baseline. More helps only if the images are consistent. Contradictory references are worse than too few.
Can I keep a character consistent without training a custom model?
Yes. Reference conditioning, character slots, and image adapters get you most of the way, especially for short projects. Training a dedicated identity model only makes sense when you have dozens of shots across many episodes and want maximum stability.
Why does the face look right in stills but drift in video?
Motion forces the model to re-synthesize geometry it cannot see. Identity decay is a function of duration, speed, and occlusion. Shorter shots, slower action, and keeping the face visible all reduce drift.
Should I generate at the highest resolution available?
Generate at the resolution the model handles best, then upscale. Very high native resolution sometimes reduces temporal stability, and upscaling with a dedicated tool often gives cleaner results.
Is it better to fix a bad shot or regenerate it?
If the failure is small — a slight eye asymmetry, a missing accessory — composite a fix. If the identity is fundamentally wrong, regenerate with a tighter reference set. Regenerating a near-miss usually produces a different near-miss.
What is the biggest time saver?
Locking the identity board and reference sheet before generating any video. Teams that skip this step spend most of their time repairing drift that a two-hour setup session would have prevented.
Where to start this week
Pick one character, one look, and one short sequence of four to six shots. Build a six-image reference pack, generate an identity board, produce a turnaround, and then animate. Run the stress tests, note where the identity fails, and fix the pack rather than the prompt. Once that loop feels routine, scale it to a full episode by keeping a reference reel for each character and reusing it as canon.
Consistency is not a feature you switch on. It is a habit of treating identity as an asset that lives outside any single generation, and every shot as a rendering of that asset. Build the asset carefully, protect it with disciplined prompting, assemble early, and repair surgically. The result is not just a character who looks the same across scenes — it is a story the audience can finally stay inside.
