How to Use Midjourney Character Reference (--cref) for Consistent AI Images

Understanding Midjourney's Character Reference Mechanics

Achieving narrative cohesion across a series of AI-generated images has historically been one of the primary obstacles for digital artists, game developers, and marketers. Midjourney solves this with the character reference parameter, denoted as --cref. This feature allows the model to extract facial features, body shapes, and key details from a reference image and project them onto new compositions. Unlike traditional image prompting, which influences the entire aesthetic, composition, and color palette of the output, character references specifically isolate and transplant the identity of the subject.

To use this feature effectively, you must understand how Midjourney parses reference images. The algorithm isolates human or humanoid elements within your source image, creating an internal map of facial ratios, hair textures, and signature details. When writing prompts, you can apply this map to new actions, environments, and emotional expressions. If you encounter issues with erratic outputs, understanding fixing issues when your AI image generator fails to output correct details can help you debug prompt interference and ratio mismatches.

To maximize your output precision, it helps to compare character reference functionality with Midjourney's other primary reference tools. The table below outlines how these parameters interact and when to deploy each one:

Pricing above reflects publicly listed rates as of August 2026. Subscription pricing changes often — confirm current rates on the provider's own pricing page before subscribing.

Step-by-Step Guide to Implementing --cref in Your Workflow

Executing consistent generations requires structured prompts and clean references. If you are unfamiliar with fundamental parameter configurations, reading our guide on mastering basic prompting structures will provide context on organizing command weights.

Step 1: Isolate a High-Quality Reference Image

The character reference algorithm performs best when the source image contains a clear, forward-facing view of the character with neutral lighting and minimal clutter. If you generated the initial character in Midjourney, upscale that specific grid quadrant and copy its direct image link. Avoid using group photos or images with busy backgrounds, as Midjourney may attempt to merge elements of the background or secondary characters into your target subject's design.

Step 2: Construct the Prompt Structure

To call the reference, construct your text prompt describing the new action or setting first, and append your parameters at the very end of the string. The syntax must look like this:

/imagine prompt: A professional software engineer sitting in a modern office, typing on a mechanical keyboard, cinematic lighting --cref https://url-to-your-image.png --cw 100 --v 6.0

Ensure that the URL is a direct link ending in an image extension such as .jpg, .jpeg, or .png. If you are using Discord, right-click the upscaled image, select "Copy Link," and paste it directly behind the --cref tag.

Step 3: Fine-Tune Fidelity with Character Weight

By default, if you omit the --cw parameter, Midjourney defaults to --cw 100. This setting carries over the face, hair, and the original clothes of your character. If your new scene requires a dramatic wardrobe change—such as placing your character in a spacesuit or formal business attire—add --cw 0. This setting isolates the face, allowing the text portion of your prompt to completely redefine the hair, clothing, and accessories.

Step 4: Combine with Style References for Aesthetic Unity

To ensure that the character sits naturally in the background without looking pasted in, combine the character reference with a style reference. Simply add --sref followed by the same URL or a different aesthetic reference URL at the end of the prompt:

/imagine prompt: A professional software engineer standing on a balcony in Tokyo, neon noir style --cref https://url-to-your-image.png --cw 0 --sref https://url-to-your-style-image.png

Advanced Multi-Character Workflows and Best Practices

Generating multiple persistent characters in a single image is a complex task because Midjourney cannot easily map separate references to separate subjects in a single generation step. To bypass this limitation, utilize the built-in pan, zoom, or Vary (Region) tools. Generate the first character in your scene using --cref, select the Region editor (inpainting tool), select the area where you want the second character to appear, change the text prompt to describe the second character, and change the --cref URL to the second character's reference image.

Additionally, avoid over-describing physical traits that are already present in your reference image. If your source image clearly shows a woman with curly red hair, writing "curly red hair" in your new prompt creates unnecessary redundancy that can cause the engine to exaggerate or distort those features. Keep the text prompt focused solely on the action, environment, and changes you wish to introduce.

Final Recommendation & Who Should Pick What

For individuals and small creative teams building visual novels, children's books, or brand mascot assets, Midjourney's --cref parameter is currently the most robust out-of-the-box solution for character persistence without requiring local hardware or deep technical knowledge of model training. If you are prioritizing speed and ease of use, leveraging Midjourney's web dashboard or Discord setup with --cref and --cw 0 offers the best balance of flexibility and rendering quality.

For enterprise-scale production requiring absolute 100% pixel-perfect consistency across complex action poses, highly specialized clothing designs, or non-human entities, consider utilizing a custom fine-tuning workflow instead. Midjourney remains excellent for prototyping, storyboarding, and direct asset generation, but highly custom production pipelines may still require hybrid setups to achieve absolute control over specialized branding assets.

Information accurate as of August 2026 — pricing and features change frequently, so verify current details on the official source before making a decision.