Midjourney --sref and --cref Explained: How to Use Image References for Style and Character Consistency

Two Parameters, Two Different Problems

Midjourney released --sref (style reference) and --cref (character reference) as separate parameters because they solve separate problems. --sref controls how the output looks: the color palette, rendering style, texture quality, lighting approach, and aesthetic feel. --cref controls who appears: the facial features, proportions, and visual identity of a specific character or person across multiple generations.

Conflating the two is the most common mistake when people first use these features. --sref will not preserve a character's face. --cref will not transfer a visual style onto a new subject. Understanding the distinction before using either parameter saves a lot of frustrated regenerations.

How --sref Works

--sref takes a URL pointing to a reference image (or multiple URLs separated by spaces) and uses that image as a style influence on the generation. The model analyzes the reference image for aesthetic properties and attempts to produce an output that has a similar visual character.

The syntax is appended to the end of your prompt: /imagine prompt: [your prompt] --sref [image URL]

The strength of the style reference is controlled by the --sw (style weight) parameter, which takes a value from 0 to 1000. The default is 100. At low values (0 to 50), the style reference has minimal influence. At high values (500 to 1000), the output closely mimics the reference aesthetic but may sacrifice prompt adherence. The practical working range for most uses is 100 to 300.

What --sref transfers well:

• Color palette and overall tonal range
• Lighting approach (flat, dramatic, ambient, backlit)
• Rendering style (painterly, photorealistic, graphic, illustrative)
• Texture density and surface quality
• Compositional mood (dark and atmospheric vs bright and clean)

What --sref does not transfer reliably:

• Specific compositional arrangements
• Specific objects or subjects in the reference image
• Fine detail at the element level (it transfers aesthetic character, not content)
• Styles that are too generic or too complex to have a distinct aesthetic fingerprint

Using Multiple --sref References

You can provide multiple reference image URLs to blend style influences: --sref URL1 URL2. You can also assign different weights to each reference by appending ::weight after each URL: --sref URL1::2 URL2::1. This gives URL1 twice the stylistic influence as URL2.

Multi-reference blending is useful when a single reference image is too specific in one direction and you want to moderate it with a second reference. For example, using one reference with the color palette you want and a second reference with the lighting quality you want, weighted to balance both influences.

The tradeoff: the more references you add, the more diluted each one becomes. Two references is usually the practical limit. Three or more often produces outputs that feel aesthetically neutral because the competing influences cancel each other out.

How --cref Works

--cref takes a reference image of a character or person and uses it to preserve visual identity across generations. The model extracts facial features, proportions, and defining visual characteristics from the reference and applies them to the generated character.

The syntax mirrors --sref: /imagine prompt: [character description and scene] --cref [image URL]

The strength is controlled by --cw (character weight), with the same 0 to 1000 scale. At --cw 0, only the character's face is referenced (body and clothing are ignored). At --cw 100 (the default), the full character appearance is used. Lowering --cw is useful when you want the character's face to carry over into a new scene but want the clothing, pose, and body proportions to be determined by the prompt rather than the reference.

What --cref transfers well:

• Facial structure and recognizable features
• Hair style and color
• Overall character proportions when --cw is at default
• Distinguishing visual features (distinctive eye shape, facial marks, etc.)

What --cref does not transfer reliably:

• Exact clothing and accessories (use --cw 100 and describe specifically in the prompt for better results)
• Very fine facial details at high variation across generations
• Realistic human likenesses with consistent photographic accuracy across many different scenes and angles
• Characters with very generic or undefined features in the source reference

Using --sref and --cref Together

Both parameters can be used in the same prompt: /imagine prompt: [description] --sref [style URL] --cref [character URL]. This generates a scene in the aesthetic of the style reference featuring a character that matches the character reference.

This combination is useful for creating consistent character content across a defined visual style — building a series of images featuring the same character in a specific aesthetic. The two parameters operate on different aspects of the generation and do not interfere with each other when used together.

The practical challenge: conflicting instructions between the two references produce inconsistent results. If the style reference features a very different rendering quality than the character reference (one is photorealistic, the other is illustrative), the model has to reconcile the conflict, and results become less predictable.

Best Practices for --sref

Use reference images that have a clear, distinct aesthetic identity. A reference image that is technically well-executed but aesthetically neutral (clean product photography, for example) produces less distinctive style transfer than a reference image with a specific visual character (a specific painting style, a film still with distinctive color grading, an illustration with a defined rendering approach).

Use your own images or images you have rights to as style references. The model extracts aesthetic properties, not content, so using a reference image stylistically does not reproduce the reference's content. However, using famous artists' work as a style reference raises ethical considerations that are worth thinking through intentionally.

Test --sw values. The default 100 is a sensible starting point but often needs adjustment for a specific reference. For strong, distinctive style references, 50 to 100 is often sufficient. For subtler references or when you want the style to dominate, 200 to 400 produces more pronounced results.

Best Practices for --cref

The quality of the reference image matters significantly. A high-resolution, well-lit, clearly defined character reference produces better consistency than a small, blurry, or ambiguous reference. The model has more to work with when the reference is clear.

Use --cw 0 when you want the character's face in a completely new scene, clothing, and pose. The face carries over while everything else is prompt-determined. This is often the most flexible approach for character consistency across varied scenes.

For illustrated or stylized characters, --cref is generally more consistent than for photorealistic characters. Stylized characters have defined, distinctive features that are easier for the model to preserve. Photorealistic likenesses at high consistency across many different scenes and angles remain a limitation of the current parameter.

Common Mistakes

Using --sref to try to preserve a character's appearance. --sref reads style, not identity. A portrait used as --sref will influence the color palette and rendering approach of the output but will not preserve the person's face. Use --cref for character identity.

Using --cref to transfer a visual style. --cref reads character identity, not aesthetic style. An illustrated scene used as --cref will attempt to preserve character-like visual properties from the reference, not the illustration style. Use --sref for style.

Setting --sw too high and losing prompt adherence. At --sw 500 and above, the style reference becomes dominant enough that the prompt description has reduced influence. Subjects, compositions, and moods from the reference begin to appear in the output. Keep --sw at 100 to 300 unless you specifically want the reference to dominate.

Using low-resolution or unclear reference images. Both parameters work better with high-quality references. Blurry, small, or compositionally complex reference images (where the relevant subject is small within a busy scene) produce less reliable results.

Expecting perfect photorealistic face consistency across many diverse scenes. --cref produces good character consistency for illustrated and stylized characters and reasonable face consistency for realistic characters in similar scenes. Expecting pixel-level identity consistency across very different angles, lighting conditions, and poses is beyond what the current parameter delivers reliably.

Quick Reference: --sref vs --cref