Midjourney Launches New Feature for Creating Consistent Characters in Multiple AI-Generated Images

The popular AI image generation service Midjourney has introduced one of its most requested features: the ability to consistently recreate characters across different images.

This capability has posed a significant challenge for AI image generators, primarily due to their reliance on diffusion models. These tools, like Stability AI's Stable Diffusion, work by processing user-provided text to create images pixel-by-pixel, using data from a vast training set of millions of human-created images.

Why Consistent Characters Matter in Generative AI

Similar to text-based large language models (LLMs) like OpenAI's ChatGPT, generative AI applications often struggle with consistency. Each prompt produces a unique output, even when the same keywords are reused. While this feature is beneficial for generating new content, it complicates situations such as storyboarding for films, novels, graphic novels, or comics—where continuity and character recognition are crucial.

To address this need, Midjourney has launched the “–cref” tag (short for “character reference”). By adding this tag to their text prompts in the Midjourney Discord, users can reference a character from a previously generated image. The AI will then aim to match the character's facial features, body type, and clothing based on the provided URL.

How to Use Midjourney's Consistent Character Feature

The “–cref” tag works best with images that have already been generated in Midjourney. Here’s a step-by-step workflow:

1. Generate or find the URL of an existing character.

2. For instance, let’s create a character with the prompt: “a muscular bald man with a beard and eye patch.”

3. Once you select your favorite image, control-click on it in the Midjourney Discord to copy the link.

4. You can then create a new prompt like, “wearing a white tuxedo standing in a villa –cref [URL],” pasting in the copied URL. Midjourney will attempt to generate the same character in your new setting.

While the outcome may not precisely match the original character or prompt, the results are promising. Users can also adjust how closely the new image resembles the original by using the “–cw” tag followed by a number between 1 and 100. A higher “cw” value means a closer match to the original character, while a lower number allows for more variation.

For example, using a low “cw” setting of 8 might yield the desired white tuxedo, but at the cost of losing distinctive features like the eyepatch—a detail that can be corrected with some tweaks.

Combining Characters

Users can create images featuring multiple characters by using two “–cref” tags with their respective URLs. This feature launched recently, and artists are already exploring its potential.

Official Notes

David Holz, Midjourney's founder, shared an update on this feature:

- Usage Instructions: Add the tag “–cref URL” after your prompt, where URL links to a character image.

- Strength Adjustments: Use “–cw” to modify reference strength, with 100 as the default. Setting it to 0 focuses solely on facial features.

- Character Focus: This feature is designed for Midjourney-created characters, not real individuals, and precision is limited—it won’t replicate minor details like dimples or logos accurately.

- Model Compatibility: The “cref” feature works for both Niji and regular Midjourney models.

Explore this innovative feature in Midjourney and start crafting more coherent narratives and visual stories!