27 мар. 2024

New Challenge for Face Recognition: Midjourney Character Reference Feature

In March 2023, Pope Francis was the talk of the world press and social media when he was spotted wearing a chic white puffer jacket from Balenciaga. It was not until after the picture was leaked to the public and went viral that it was revealed to have been created by the Midjourney neural network. In a comment on its AI provenance, Reuters said that the image “sowed some confusion online” since viewers mistakenly believed it to be real. Source: Image of Pope Francis wearing oversized white puffer coat is AI-generated.

The fake photo generated by an AI tool that creates images from short text prompts was really intriguing. People were taken aback by the unexpected image of the Pope wearing a winter coat, contrasting with the traditional photos of Francis in his usual attire for papal portraits.

The photo of the pontiff generated interest due to its humorous nature. However, it is important to approach AI-generated content with critical thinking and ethical considerations, especially when it involves depictions of public figures or religious leaders.

Read this blog post to learn how Midjourney and deepfake detection are connected and how liveness detection and face recognition technologies are used to combat deepfake.

How The Midjourney Character Reference Feature Works

Midjourney neural network was initially developed for artists and creatives to help them generate and manipulate visual content. However, as the technology evolved, it became clear that the neural network had potential applications beyond the realm of art. Its ability to analyze and interpret large amounts of data quickly and efficiently made it useful for a wide range of industries, including finance, healthcare, and marketing. Today, Midjourney neural network is used in various fields to improve decision-making, automate processes, and generate insights from complex data sets. Its versatility and adaptability have made it a valuable tool for professionals in diverse fields.

At the same time, the network can potentially be used for malicious purposes, including creating deepfake content like realistic-looking but false images. As a result, Midjourney can be misused to create and spread misleading or harmful content, such as fake news, scams, or defamation.

The Midjourney Character Reference Feature (Cref) is a new feature used to create realistic and accurate character references and place them in different contexts. First, the user creates a character in the Midjourney network. After that the generated image can be transferred to different settings using the Cref feature. Then the prompt is written as usual, but at the end the user has to add “-cref (URL of the basic image)-cw 100.” Cw stands for “character weight” and has a scale from zero to 100. Closer to zero will give you more character variation, while 100 will keep the character image unchanged.

The ability to generate the varying character images makes Cref a tool for deepfake generation as it helps to fill in the gaps in the image data base of the created false image. In other words, you can obtain a person’s image that isn’t present on social media or other information sources.

This feature can be used for malicious purposes in the context of deepfake by malicious actors to generate fake character references to impersonate individuals and spread false information or propaganda. For example, a malicious actor could use the Midjourney Character Reference Feature to create a deepfake video of a politician saying or doing something they never actually did, in order to manipulate public opinion or damage their reputation. This could potentially have serious real-world consequences, such as influencing elections or inciting violence.

Additionally, the feature could be used to create fake character references of individuals in compromising or illegal situations, leading to false accusations or blackmail.

Overall, the Midjourney Character Reference Feature could be exploited for harmful and deceptive purposes in the context of deepfake technology, undermining trust and creating chaos in various spheres of society.

The Limitations of Midjourney-Generated Images

Midjourney images, while high in quality, have some drawbacks that can potentially make them detectable. Some of these drawbacks include the lack of fine detail and consistency in AI-generated faces, as well as a focus on artistic style over realism. These factors can lead to inconsistencies or abnormalities in the images that could potentially be flagged as artificial.

Midjourney-generated images have the following limitations:

  • AI-generated faces lack detail and consistency. Compared to real photos, faces in Midjourney images might have unusual features, unrealistic textures, or variations across different outputs for the same prompt.

  • Focus on artistic style over realism. Midjourney prioritizes creating an image that matches your artistic prompt, so facial details might be stylized or exaggerated. Midjourney-generated images may lack the realism and authenticity that is captured in real-life photographs, making it difficult to accurately depict real-world scenarios.

  • Limited resolution. The images generated during midjourney may have a limited resolution, which may result in a lack of clarity and detail in the images.

  • Limited diversity. The variety of scenes and objects that can be generated during midjourney may be limited, resulting in a lack of diversity in the generated images.

  • Inability to capture dynamic changes. Midjourney-generated images may not be able to accurately capture dynamic changes in the environment, such as movement or changes in lighting conditions.

  • Lack of contextual understanding. Midjourney-generated images may not have a deep understanding of the context and surroundings, resulting in inaccuracies in the depiction of scenes.

  • Inaccuracy in depiction. Due to the limitations in understanding context and surroundings, midjourney-generated images may be inaccurate in the representation of real-world objects and scenes.

As technology continues to advance, it is essential to address these limitations and prevent the potential misuse of AI technologies by developing safeguards and regulations.

Face Recognition and Midjourney-Generated Faces

Traditional face recognition systems are not designed to detect Midjourney-generated faces since face recognition aims to identify known individuals, not determine if a face is real or AI-generated. Plus, it works by comparing facial features to a database of real faces. Midjourney faces, while realistic, may not have exact matches in the database.

However, there are some deepfake detection solutions that go beyond traditional face recognition and use more complex technology:

AI analysis. These tools use advanced algorithms to analyze things like blinking patterns, subtle facial movements, and even skin texture to identify inconsistencies that might point to an AI-generated face.

Training on AI-generated data. Some deepfake detectors are specifically trained on datasets of AI-generated faces, including those from Midjourney. This can improve their ability to spot these types of faces.

Deepfake Detection Solutions and Midjourney-Generated Faces

Deepfakes are a type of synthetic media that uses artificial intelligence to create realistic and convincing videos or images. They can be used to mimic individuals, celebrities, or public figures, causing ethical, legal, and security concerns.

Here are some of the key points about deepfakes:

  • How they are created. Deepfakes are created using deep learning techniques and algorithms. This involves collecting a large amount of training data, preprocessing the data, using generative adversarial networks (GANs) to train the model, and then face-swapping the target person's face with another person's face.

  • Challenges of detection. Deepfakes are becoming increasingly sophisticated, making them difficult to detect. Additionally, data scarcity and the adversarial nature of deepfake technology pose challenges for detection systems.

  • Facial recognition and deepfakes. Facial recognition technology can be used to help detect deepfakes, but it has limitations. For example, facial recognition often relies on analyzing static features, which deepfakes can mimic. Additionally, facial recognition systems typically perform best with head-on views and may not be able to detect deepfakes with profile shots or extreme angles.

We explored the creation of deepfakes, the latest deepfake detection methods, and the vulnerabilities of face recognition systems in more detail here: Facial Recognition for Deepfake Detection.


Midjourney is currently evolving quickly as the developers strive to meet the changing needs of their customers in the digital age. They are embracing new technologies and digital marketing strategies to stay ahead of the competition and provide a seamless and engaging experience for their customers.

Artists may find the new Midjourney Character Reference Feature helpful in producing characters fast. However, this presents a problem for facial recognition technology as attackers may also use features like Cref to generate the same person from different positions to bypass liveness detection.