BackBlog
General

How to Build a Master Reference Pack for AI Video Character Consistency

A master character reference pack locks in facial features, style, and identity across every AI video generation mode, eliminating face drift and style inconsistency without rework. This tutorial walks you through building a 4-asset reference pack and deploying it across text-to-video, image-to-video, and motion-transfer workflows in Seedance and Kling. You’ll get reusable prompt templates, model-specific settings, and a face-swap fallback for any residual drift.

How to Build a Master Reference Pack for AI Video Character Consistency
MagicBoat AI7 min read34 views
Table of contents

Why AI Video Character Consistency Breaks

You generate a perfect character portrait, animate it into a strong opening shot, and then discover that the next angle has a wider nose, different hair, or a shifted skin tone. These failures are usually caused by unanchored generation rather than bad prompting. Every text-to-video model interprets character descriptions differently, so even small prompt changes can produce face morphing, mismatched features, style drift, or lost details such as scars and freckles.

A master character reference pack creates a single source of truth for the character’s identity. Instead of describing the character from scratch for every shot, you reuse tested assets that define the face, body proportions, distinguishing features, and visual style.

The Four Assets in a Master Reference Pack

  1. A high-quality master reference portrait
  2. A reusable identity prompt block
  3. Saved model-specific seeds and generation settings
  4. A face-swap source image for correcting residual drift

Three Ways to Lock Character Identity

  • Visual reference anchoring: Use a master portrait as the primary identity reference for image-to-video and motion-transfer workflows.
  • Prompt-based identity locking: Start every text-to-video prompt with the same structured description of permanent character traits.
  • Post-generation correction: Apply face swap when fast movement, close-ups, or difficult camera angles cause visible drift.

Step 1 — Build the Master Reference Portrait

The master portrait is the most important asset in the pack. It should clearly display the character’s defining features without variable elements that could confuse later generations.

  • Use a three-quarter front view to show both frontal and profile details.
  • Choose a neutral, relaxed expression.
  • Use even, diffused lighting without deep shadows.
  • Select a plain white, gray, or beige background.
  • Generate at a minimum width of 1024 pixels with sharp facial details.

Generate the portrait with an image model such as Flux or Seedream and save the seed when you find a result you like.

photorealistic 3/4 portrait of a 28-year-old woman, wavy auburn shoulder-length hair, hazel eyes with green flecks, freckles across nose and cheeks, soft neutral expression, plain light gray background, even studio lighting, sharp focus, high detail --ar 2:3 --seed 123456

Expected result: One clear portrait that captures every fixed identity trait without distracting clothing, accessories, or scenery.

Save the final image together with its exact prompt, seed, model name, and generation settings. You can reuse these details when creating additional reference angles.

Step 2 — Write a Reusable Identity Prompt Block

Create a fixed 30–50 word description containing only permanent traits. Paste this block at the beginning of every prompt for the character.

Fixed identity traitsVariable scene traits
Face shape, jawline, and nose shapeClothing, shoes, and accessories
Eye color, hair color, and hair textureExpression, pose, and action
Skin tone, scars, freckles, or birthmarksBackground, setting, and lighting
Body type, height, and age rangeCamera angle, shot size, and effects

Example identity block:

28-year-old woman with a heart-shaped face, small straight nose, hazel eyes with green flecks, wavy auburn shoulder-length hair, freckles across her nose and cheeks, fair skin, and a slim 5-foot-6-inch build

Add variable scene information after the identity block:

[IDENTITY BLOCK], wearing a navy blue sweater and jeans, walking slowly down a sunlit city sidewalk, hands in pockets, casual gait, medium full shot, natural daylight, cinematic --motion 0.6 --style 0.85

Expected result: A copy-and-paste identity description that remains unchanged across prompts while clothing, action, setting, and camera details vary.

Step 3 — Save Model-Specific Settings

Video models interpret prompts and references differently. Seedance may favor softer facial details and stylized movement, while Kling may preserve sharper features and proportions. Test each model and record the settings that most closely reproduce the master portrait.

SettingSeedance starting pointKling starting point
Style strength for text-to-video0.8–0.90.7–0.8
Motion strength for close-ups0.5–0.70.4–0.6
Motion strength for full-body shots0.6–0.80.5–0.7
Reference weight for image-to-video0.7–0.80.8–0.9
Seed usageLock for text-to-videoLock across modes when possible

These values are starting points rather than guarantees. Available controls and their behavior can change by model version or platform implementation.

Before producing long clips, generate a two-second test in each model and compare the face, hair, body proportions, and overall style with the master reference.

Step 4 — Deploy the Pack Across Generation Modes

Text-to-Video

Place the identity block at the beginning of the prompt, followed by the outfit, action, environment, camera direction, and style. Apply the saved seed and model-specific style settings. This mode is useful for establishing shots and scenes without a required starting pose.

Image-to-Video

Use the master portrait or a pose-adjusted derivative as the input image. Repeat the identity block in the prompt and apply the saved motion and reference-weight settings. This is generally the most reliable mode for close-ups and medium shots where facial accuracy matters.

Motion Transfer

Upload the character reference and a source motion clip. Enable identity-preservation controls when available and include the identity block in the prompt. Motion transfer is useful for dance, action, and other complex movement that is difficult to specify with text alone.

Step 5 — Audit and Correct Residual Drift

Review every generated clip against the master portrait. Check face shape, eye spacing and color, nose shape, hairline, freckles or scars, skin tone, and body proportions. Pay special attention to frames with fast head movement, profile views, partial occlusion, or extreme expressions.

If a strong clip contains visible facial drift, use face swap with the same master portrait rather than regenerating the entire performance.

  • Correct only clips where the face is visible enough for drift to matter.
  • Use the same master portrait as the source identity for every correction.
  • Start with an identity strength around 0.7–0.8, if that control is available.
  • Inspect the result frame by frame for alignment errors, flicker, or lost expressions.

Face swap should preserve the original motion and timing, but results depend on visibility, head angle, motion blur, and the specific tool.

Advanced Tips for Long-Form Projects

  • Add more angles: Create profile, full-body, and three-quarter rear references when the shot list requires them.
  • Create outfit sub-packs: Save one full-body reference for each major outfit while keeping the core identity block unchanged.
  • Separate visual styles: Maintain different packs for photorealistic, illustrated, and animated versions of the character.
  • Version the pack: Assign version numbers whenever the character design changes so old and new references are not mixed.
  • Keep a continuity log: Record the model version, seed, prompt, aspect ratio, reference weight, motion strength, and corrections used for every approved clip.

Limitations

  • Extreme close-ups may reveal minor skin-texture differences between models.
  • Unusual facial hair, prosthetics, or highly distinctive anatomy may require several references.
  • Fast or erratic head movement can cause both generation drift and face-swap alignment errors.
  • A character may not transfer cleanly between photorealistic and animated styles without separate reference packs.
  • Seeds and similarly named controls do not necessarily behave the same way across models or model updates.

FAQ

Can the same reference pack be used for Seedance and Kling?

Yes. Reuse the same identity assets, but test and save separate settings for each model. Generate short comparison clips before committing to longer shots.

How many reference images are required?

One strong three-quarter portrait may be enough for a short project. Add a full-body image for clothing and proportion continuity, plus a profile image for side-view shots.

Can the character’s outfit change?

Yes. Keep the identity block and face reference unchanged, then modify only the clothing description. For recurring outfits, create outfit-specific full-body references.

Is the same seed necessary for every clip?

A fixed seed can reduce variation in text-to-video workflows. In image-to-video and motion transfer, the reference image is usually the stronger identity anchor, although seed behavior varies by model.

Will face swap preserve motion and expression?

It is designed to preserve the underlying performance, but the result should still be checked for flicker, alignment errors, and expression loss.

Final Checklist

  • Master reference portrait with a three-quarter view, neutral lighting, and plain background
  • Reusable identity prompt block containing only fixed traits
  • Saved seeds, style strengths, motion strengths, and reference weights for each model
  • Consistent source portrait for face-swap corrections
  • Additional angle or outfit references when required
  • Continuity log for every approved clip