Models - Mar 11, 2026

How to Maintain Consistent Characters in Midjourney V7: A Masterclass

How to Maintain Consistent Characters in Midjourney V7: A Masterclass

Introduction

Character consistency has been the white whale of AI image generation. Generating a single beautiful character is easy. Generating the same character in different poses, lighting conditions, environments, and emotional states — while maintaining recognizable identity — has been extraordinarily difficult.

When Midjourney V7 launched on April 4, 2025, its character reference system represented the most significant progress on this problem to date. The --cref parameter, combined with V7’s improved understanding of human features, enables workflows that were previously impossible: storyboarding, sequential art, brand character development, children’s book illustration, and marketing campaigns with consistent character representation.

This article is a practical masterclass. It covers the mechanics, the techniques, the limitations, and the workflows that make character consistency work in V7.

Understanding Character Reference

How --cref Works

The character reference parameter tells V7 to use one or more reference images as the basis for character appearance in a new generation. The syntax is straightforward:

/imagine a woman standing in a rainy city street, film noir lighting --cref [image URL]

The model analyzes the reference image to extract character features — face structure, hair color and style, body type, skin tone, and distinguishing characteristics. It then applies these features to the new generation while following the rest of the prompt.

What Gets Preserved

V7’s character reference preserves:

  • Facial structure: Bone structure, eye shape, nose profile, lip shape, jawline
  • Coloring: Skin tone, eye color, hair color
  • Hair: Style, length, texture (with some variation depending on the prompt context)
  • Body type: General proportions and build
  • Distinguishing features: Freckles, scars, glasses, facial hair, and other identifying details

What Changes

The reference system intentionally allows variation in:

  • Expression: The character can smile, frown, look surprised, or show any emotion
  • Clothing: Unless specifically prompted to match, clothing will change based on the scene description
  • Pose: The character can be in any position or action
  • Aging: Subtle age variations may occur depending on the context
  • Lighting and color grading: The character will be lit according to the scene, not the reference

This flexibility is by design. A character reference system that rigidly copied every detail would be useless for storytelling — characters need to express different emotions, wear different clothes, and exist in different environments while remaining recognizably themselves.

Building a Strong Reference

The Initial Generation

The quality of your character reference determines the consistency of everything that follows. Invest time in generating the ideal reference image before starting a multi-image project.

Best practices for reference images:

  1. Clear face visibility: Front-facing or three-quarter views with unobstructed facial features work best. Profile views or partially obscured faces provide less information for the model to work with.

  2. Neutral expression: Start with a neutral or mild expression. Extreme expressions (wide smile, screaming, crying) can bias the model toward reproducing that expression in subsequent generations.

  3. Good lighting: Even, well-lit reference images give the model the clearest data about the character’s features. Dramatic shadows or unusual color lighting can confuse the extraction.

  4. Appropriate resolution: Use V7’s upscaler to create a high-resolution reference. More pixel data means more feature information for the model.

  5. Distinctive features: If your character has distinguishing features — a specific hairstyle, glasses, a scar, a tattoo — make sure these are clearly visible in the reference.

Using Multiple References

V7 supports multiple character references, which can improve consistency:

/imagine a warrior standing on a cliff at sunrise --cref [URL1] [URL2] [URL3]

Using 2-3 references of the same character from different angles gives the model a more complete understanding of the character’s three-dimensional appearance. This is particularly effective when references show:

  • Front view
  • Three-quarter view
  • Side view

The Character Weight Parameter

The --cw parameter controls how strongly the character reference influences the generation:

  • --cw 100 (default): Strong character adherence. The generated character will closely match the reference.
  • --cw 50: Moderate adherence. The character will be recognizably similar but with more variation.
  • --cw 0: Face only. Only facial features are referenced; body type, hair, and other features are generated freely.

For most consistency workflows, --cw 100 is appropriate. Lower values are useful when you want to maintain face identity but change other aspects — for example, showing a character at different ages, with different hairstyles, or after a dramatic transformation.

Workflow Techniques

The Reference Sheet Method

Professional character designers often create reference sheets — a single image showing a character from multiple angles with consistent design. V7 can generate these directly:

/imagine character reference sheet, young woman with short red hair, green eyes, freckles, athletic build, front view, side view, back view, three-quarter view, white background, character design --ar 16:9

The resulting reference sheet can then be used as the --cref source for all subsequent generations. This approach provides the model with comprehensive visual data about the character from a single, consistent source.

Scene Progression Workflow

For storytelling projects (comics, storyboards, children’s books), the workflow is:

  1. Generate the definitive character reference — invest time in getting this right
  2. Plan your scenes — write prompts for each scene before generating
  3. Generate scenes sequentially — using the same --cref for every generation
  4. Review consistency — compare each generation against the reference and previous scenes
  5. Regenerate outliers — some generations will drift from the reference; regenerate these
  6. Refine with inpainting — use V7’s web editor to fix minor inconsistencies in specific regions

Multi-Character Scenes

When a scene includes multiple established characters, V7 allows multiple character references with positional guidance:

/imagine two people having coffee at a café, the man on the left is talking, the woman on the right is laughing --cref [man URL] [woman URL]

This is less reliable than single-character references. V7 may occasionally swap which reference applies to which figure, or blend features between characters. Techniques to improve multi-character consistency:

  • Distinct character designs: Characters with very different appearances (different hair colors, body types, ethnicities) are easier for the model to keep separate
  • Clear positional language: “on the left,” “on the right,” “in the foreground,” “behind” help the model assign references correctly
  • Generate and iterate: Multi-character consistency often requires multiple generations to get right

Style Reference Combination

Character references can be combined with style references for comprehensive visual consistency:

/imagine a detective examining clues in a dimly lit office --cref [character URL] --sref [noir style URL] --sw 50

This produces output that is consistent in both character appearance and visual style — essential for projects like graphic novels, marketing campaigns, or branded content series.

Common Problems and Solutions

The Drift Problem

Over many generations, character appearance can gradually drift from the reference. Each generation introduces small variations, and if you use output images as references for subsequent generations, these variations accumulate.

Solution: Always reference back to the original reference image, not to recent generations. Keep your definitive character reference saved and use it consistently.

The Expression Lock

Sometimes the model reproduces the reference’s expression even when the prompt requests a different emotion. This is more common with highly expressive reference images.

Solution: Use a reference with a neutral expression. If you need a specific expression, be explicit in the prompt: “wide smile showing teeth,” “furious expression with furrowed brows,” “gentle, melancholic gaze.”

The Clothing Carryover

The model may reproduce the reference’s clothing even when the prompt describes different attire. This is more likely with distinctive or prominent clothing in the reference.

Solution: Be specific about clothing in the prompt. “Wearing a blue business suit” is clearer than “in formal attire.” If clothing carryover persists, use --cw 0 to reference face only and describe all other features in the prompt.

The Angle Problem

References from certain angles may not translate well to dramatically different viewpoints. A front-facing reference may produce inconsistent results when the prompt requests a rear view.

Solution: Use multiple reference angles. Include at least a front and three-quarter reference for best results across all viewpoints.

Advanced Techniques

Age Progression

To show a character at different ages while maintaining identity:

  1. Generate the character at their “base” age with a strong reference
  2. For older versions: prompt with age-specific details (“same person at 60, grey hair, wrinkles”) with --cw 50 to allow aging while maintaining core features
  3. For younger versions: prompt with youth-specific details (“same person as a teenager”) with similar reduced character weight

Stylistic Adaptation

To maintain character identity across different artistic styles:

  1. Establish the character in a realistic style first
  2. Use --cref with style-specific prompts: “in the style of a Pixar character,” “as a watercolor illustration,” “in anime style”
  3. The model will adapt the character’s appearance to the style while maintaining identifying features

Niji 7, Midjourney’s anime-focused model released in January 2026, works with character references to create anime versions of characters established in V7’s realistic style. The cross-model consistency isn’t perfect but is often surprisingly effective.

Brand Character Development

For marketing and branding use cases:

  1. Generate the brand character with specific brand-relevant features
  2. Create a comprehensive reference sheet
  3. Generate the character in all needed contexts: product shots, lifestyle scenes, social media formats, print advertising
  4. Use style references to maintain brand visual consistency across all assets
  5. Build a library of approved generations to use as additional references

Limitations to Acknowledge

Not Perfect Consistency

V7’s character reference system is remarkable but not flawless. Across many generations, you will encounter variations. The system works on a spectrum from “very consistent” to “recognizably the same person” — it does not produce pixel-identical characters.

For applications requiring absolute consistency (animation, where characters must be identical frame to frame), V7’s character reference is a starting point that requires additional work — either manual editing or post-processing.

No Public API

Midjourney’s lack of a public API means that character consistency workflows cannot be automated. Every generation requires manual interaction through the web interface or Discord. For large-scale projects requiring hundreds of consistent character images, this is a significant bottleneck.

The ongoing copyright lawsuits from Disney/Universal (June 2025) and Warner Bros. (September 2025) add legal uncertainty to commercial use of Midjourney-generated characters. For brand characters that will be used in major commercial contexts, this risk should be evaluated.

Adobe Firefly offers commercially safe generation but doesn’t have comparable character consistency features. Flux, the open-source alternative, can achieve character consistency through LoRA fine-tuning, but this requires significantly more technical expertise.

The Creative Opportunity

Character consistency in AI generation opens creative doors that didn’t exist a year ago. Independent creators can now produce visual narratives — comics, illustrated stories, storyboards, character-driven marketing — without the cost of hiring illustrators for every variation of every character in every scene.

This doesn’t replace illustrators. It changes the economics and speed of certain types of visual creation. A concept that would have required dozens of commissioned illustrations can now be prototyped in hours, tested with audiences, and refined before any traditional illustration begins.

For creators working across multiple AI tools — generating characters in Midjourney, writing narratives in text models, planning projects in productivity tools — the challenge becomes managing the workflow across platforms. AI workspace tools like Flowith help by providing a unified environment where creative ideation, generation, and organization can happen without constantly switching between specialized tools.

References