Describe the vibe, see the look: An AI-based system projects makeup onto the user’s face

Combining artificial intelligence and projection technology to create realistic, personalized makeup previews

Institute of Science Tokyo

FacebookXLinkedInWeChatBlueskyMessageWhatsAppEmail

 

image: 

The system interprets mood-based phrases to generate makeup palettes and projects them onto the user’s face, giving a realistic preview of how the makeup would appear under natural lighting and on the person’s own complexion and skin texture.

view more 

Credit: Institute of Science Tokyo

An artificial intelligence-based projection makeup system from Science Tokyo lets users describe a mood or style in their own words and instantly see matching makeup colors on their faces. The technology learns each person’s preferences in real time and displays results under realistic lighting that reflects individual skin tone and texture, making it more true to life than traditional virtual makeup apps that project effects onto two-dimensional displays.

Finding the right makeup color is an important part of the user experience when shopping for cosmetics. Virtual makeup technologies, which typically use augmented reality to overlay makeup effects through a smartphone or tablet, have made experimenting easier. However, choosing the right colors from hundreds of options can still feel overwhelming, and the results often appear artificial on a flat screen, failing to mimic how makeup appears on real skin under natural light.

Now, researchers from Institute of Science Tokyo (Science Tokyo), Japan, have created a new system that turns a user’s spoken impressions that describe a mood or theme, with phrases like ‘Sakura in spring,’ directly into personalized makeup colors. This method combines an image generation artificial intelligence (AI) model with a projection system that lets users see the makeup simulated on their actual faces in real life, while the generated colors are refined using real cosmetic color distributions.

The study was led by graduate student Kemeng Zhang, graduate student Hao-Lun Peng, and Associate Professor Yoshihiro Watanabe from the Department of Information and Communications Engineering, Science Tokyo. The results were published online in the International Journal of Human-Computer Interaction on January 21, 2026. A demonstration video illustrating the interactive personalized color exploration process is available on YouTube.

“Users can easily explore preferred makeup colors from a large number of combinations through interactive optimization using impression words and projection-based makeup. This can help non-expert users efficiently find satisfying results in the vast space of color combinations,” says Watanabe.

In this impression-guided text-to-makeup-color model, users simply describe the vibe they want, and the system translates it into makeup color suggestions. Users are encouraged to imagine scenes, objects, or moods and describe them naturally, using phrases such as ‘night rose’ or ‘autumn forest with warm sunlight.’ The AI then generates a reference image representing that impression and produces five suggested color themes for the cheeks, eyeshadow, and lips.

These colors are projected directly onto the user’s face using a high-speed dynamic projection mapping setup. This system uses a high-speed projector and camera to ensure that the makeup remains correctly aligned with the user’s face even as they move. The system tracks the user’s facial features in real time, and users can view the results in a mirror.

This system accounts for how different skin tones and lip colors reflect light, making the simulation more realistic than viewing makeup on a two-dimensional screen. As users select their preferred options from the projected results, the system updates its suggestions using an optimization method that gradually learns their preferences.

“With such a system, users can simply describe their desired impression of makeup colors in natural language and observe the effects in the mirror,” says Watanabe.

The system showed strong performance in user evaluations. In an online survey of a hundred participants, users reported that the system produced appropriate color suggestions for their impression texts. In a hands-on study with fifteen users comparing the system to a manual color adjustment tool, participants found the impression-guided method faster and more intuitive. Many said they enjoyed being able to quickly try a wide range of makeup styles, and some discovered appealing color combinations they would not have chosen on their own.

The system was also rated highly by experts from the cosmetics industry, who noted its potential for both everyday users and professionals, such as exploring makeup ideas for themed fashion shows or supporting early-stage product development by quickly generating color concepts from abstract design ideas.

This makeup generation and recommendation system highlights the growing role of AI in creative fields, helping consumers explore makeup styles they enjoy and offering beauty professionals new ways to develop and test ideas.

 

***

 

About Institute of Science Tokyo (Science Tokyo)
Institute of Science Tokyo (Science Tokyo) was established on October 1, 2024, following the merger between Tokyo Medical and Dental University (TMDU) and Tokyo Institute of Technology (Tokyo Tech), with the mission of “Advancing science and human wellbeing to create value for and with society.”

---

Journal

International Journal of Human-Computer Interaction

DOI

10.1080/10447318.2025.2599521 

Method of Research

Computational simulation/modeling

Subject of Research

People

Article Title

Impression-Guided Interactive Personalized Color Exploration Framework for Dynamic Projection Mapping Makeup

 

AI’s game-playing still has flaws: AlphaZero-style self-play tested on Nim

Despite heavy training, agents show blind spots and can miss optimal moves

Queen Mary University of London

FacebookXLinkedInWeChatBlueskyMessageWhatsAppEmail

 

image: 

An illustration of Nim game states. The left panel shows an initial board configuration of five heaps: [n1, n2, n3, n4, n5] = [1, 3, 5, 7, 9]. The middle panel depicts an intermediate board state during gameplay: [v1, v2, v3, v4, v5] = [1, 2, 4, 4, 3], resulting from players removing counters. The right panel represents the terminal state where all counters are cleared, signifying a win for the player who made the last move.

view more 

Credit: Image by Dr Bei Zhou, Research Associate at Imperial College, London, and Dr Søren Riis Reader in Computer Science Queen Mary University of London.

• 
  

• Researchers tested AlphaZero-style self-play on Nim, a children’s game solved mathematically. 
• Despite heavy training, agents show blind spots and can miss optimal moves. 
• Strong performance does not always mean the system has learned the underlying winning rule. 

Embargo: immediate.

New research published in Machine Learning shows pattern learning is not enough to train AI to tackle games – and abstract representations or hybrid approaches may help.  

Many AI researchers describe game-playing as the “Formula 1” of AI: it’s a controlled test environment with clear rules and clear success criteria. This paper uses that idea as a diagnostic, by studying a very simple game Nim, a children’s matchstick game whose optimal strategy is known exactly.  

Because the correct move is known for every position, we can measure whether an agent plays optimally across the state space. The research found that while small boards can work, despite heavy training and search, agents show blind spots and can miss optimal moves, and performance degrades as the board grows, with predictions approaching random. This suggests impartial games often need analytic representations, not pattern learning. 

What does this mean for gaming with machines? 
Self-play AIs can be very strong, but in games where both players share the “pieces” and the winning strategy is an abstract arithmetic rule, pattern-recognition from raw positions may not be enough on its own.  

Wider implications: 
The results don’t diminish the achievements of self-play AI in games like chess and Go. Rather, they help map where today’s methods can struggle, and where more abstract representations or hybrid approaches may be beneficial. More broadly, it’s a reminder that systems can perform well in common cases while remaining brittle in rare-but-important ones. 

Dr Søren Riis Reader in Computer Science at Queen Mary University of London said: “Nim is a children’s game with a complete mathematical solution, yet AlphaZero-style self-play can still develop blind spots—becoming competitive while missing optimal moves across many positions.” 

“This suggests that, for future work in AI, impressive performance alone is not proof that a system has learned the underlying principle: methods that capture abstract structure may be needed to reduce blind spots.” 

-- 

 “Impartial Games: A Challenge for Reinforcement Learning” by Dr Bei Zhou Research Associate at Imperial College London, and Dr Søren Riis Reader in Computer Science at Queen Mary University of London is published in Machine Learning 

 

---

Journal

Machine Learning

DOI

10.1007/s10994-026-06996-1 

Method of Research

Experimental study

Subject of Research

People

Article Title

Impartial Games: A Challenge for Reinforcement Learning

Article Publication Date

13-Mar-2026

Improving AI models’ ability to explain their predictions

A new approach could help users know whether to trust a model’s predictions in safety-critical applications like health care and autonomous driving.

Massachusetts Institute of Technology

FacebookXLinkedInWeChatBlueskyMessageWhatsAppEmail

Cambridge, MA — In high-stakes settings like medical diagnostics, users often want to know what led a computer vision model to make a certain prediction, so they can determine whether to trust its output.

Concept bottleneck modeling is one method that enables artificial intelligence systems to explain their decision-making process. These methods force a deep-learning model to use a set of concepts, which can be understood by humans, to make a prediction. In new research, MIT computer scientists developed a method that coaxes the model to achieve better accuracy and clearer, more concise explanations.

The concepts the model uses are usually defined in advance by human experts. For instance, a clinician could suggest the use of concepts like “clustered brown dots” and “variegated pigmentation” to predict that a medical image shows melanoma.

But previously defined concepts could be irrelevant or lack sufficient detail for a specific task, reducing the model’s accuracy. The new method extracts concepts the model has already learned while it was trained to perform that particular task, and forces the model to use those, producing better explanations than standard concept bottleneck models.

The approach utilizes a pair of specialized machine-learning models that automatically extract knowledge from a target model and translate it into plain-language concepts. In the end, their technique can convert any pretrained computer vision model into one that can use concepts to explain its reasoning.

“In a sense, we want to be able to read the minds of these computer vision models. A concept bottleneck model is one way for users to tell what the model is thinking and why it made a certain prediction. Because our method uses better concepts, it can lead to higher accuracy and ultimately improve the accountability of black-box AI models,” says lead author Antonio De Santis, a graduate student at Polytechnic University of Milan who completed this research while a visiting graduate student in the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT.

He is joined on a paper about the work by Schrasing Tong SM ’20, PhD ’26; Marco Brambilla, professor of computer science and engineering at Polytechnic University of Milan; and senior author Lalana Kagal, a principal research scientist in CSAIL. The research will be presented at the International Conference on Learning Representations.

Building a better bottleneck

Concept bottleneck models (CBMs) are a popular approach for improving AI explainability. These techniques add an intermediate step by forcing a computer vision model to predict the concepts present in an image, then use those concepts to make a final prediction.

This intermediate step, or “bottleneck,” helps users understand the model’s reasoning.

For example, a model that identifies bird species could select concepts like “yellow legs” and “blue wings” before predicting a barn swallow.

But because these concepts are often generated in advance by humans or large language models (LLMs), they might not fit the specific task. In addition, even if given a set of pre-defined concepts, the model sometimes utilizes undesirable learned information anyway, which is a problem known as information leakage.

“These models are trained to maximize performance, so the model might secretly use concepts we are unaware of,” De Santis explains.

The MIT researchers had a different idea: Since the model has been trained on a vast amount of data, it may have learned the concepts needed to generate accurate predictions for the particular task at hand. They sought to build a CBM by extracting this existing knowledge and converting it into text a human can understand.

In the first step of their method, a specialized deep-learning model called a sparse autoencoder selectively takes the most relevant features the model learned and reconstructs them into a handful of concepts. Then, a multimodal LLM describes each concept in plain language.

This multimodal LLM also annotates images in the dataset by identifying which concepts are present and absent in each image. The researchers use this annotated dataset to train a concept bottleneck module to recognize the concepts.

They incorporate this module into the target model, forcing it to make predictions using only the set of learned concepts the researchers extracted.

Controlling the concepts

They overcame many challenges as they developed this method, from ensuring the LLM annotated concepts correctly to determining whether the sparse autoencoder had identified human-understandable concepts.

To prevent the model from using unknown or unwanted concepts, they restrict it to use only five concepts for each prediction. This also forces the model to choose the most relevant concepts and makes the explanations more understandable.

When they compared their approach to state-of-the-art CBMs on tasks like predicting bird species and identifying skin lesions in medical images, their method achieved the highest accuracy while providing more precise explanations.

Their approach also generated concepts that were more applicable to the images in the dataset. 

“We’ve shown that extracting concepts from the original model can outperform other CBMs, but there is still a tradeoff between interpretability and accuracy that needs to be addressed. Black-box models that are not interpretable still outperform ours,” De Santis says.

In the future, the researchers want to study potential solutions to the information leakage problem, perhaps by adding additional concept bottleneck modules so unwanted concepts can’t leak through. They also plan to scale up their method by using a larger multimodal LLM to annotate a bigger training dataset, which could boost performance.

###

This research was supported by the Progetto Rocca Doctoral Fellowship, the Italian Ministry of University and Research under the National Recovery and Resilience Plan, Thales Alenia Space, and the European Union under the NextGenerationEU project.

US likely used AI in airstrike that killed 160 schoolgirls: Report

March 11, 2026 

A view of the debris of a school, where many students and teachers lost their lives on the first day of the wave of attacks launched by the United States and Israel against Iran, in Hormozgan, Iran on March 05, 2026. [Stringer – Anadolu Agency]

Artificial intelligence helped the US identify targets in the opening phase of Washington’s war against Iran, British newspaper The Times reported Wednesday, raising questions about a strike that killed more than 175 students and staff at a girls school, Anadolu reports.

The newspaper reported that in the first 24 hours of Operation Epic Fury, US forces struck more than 1,000 targets in Iran with the assistance of AI systems designed to analyze large volumes of intelligence data and suggest potential strike locations.

The pace, about 42 suggested targets per hour, has led analysts to question whether the speed of automated systems may be outstripping the ability of humans to fully verify targets.

The scrutiny follows a strike on the Shajareh Tayyebeh primary school in the southern Iranian city of Minab that killed at least 160 school girls and others.

Evidence examined by analysts suggests the site was likely hit by US Tomahawk cruise missiles.

Historical satellite imagery showed the school had once been part of an Islamic Revolutionary Guard Corps (IRGC) complex but had been separated from it for nearly a decade.

Experts told the newspaper that outdated imagery or automated analysis could have contributed to the site being identified as a military target. The Pentagon said an investigation into the strike is ongoing.

The Times said the US and Israel are deploying several artificial intelligence systems in the conflict, including tools developed under the Pentagon’s Project Maven program, which uses machine learning to analyze intelligence gathered from satellites, surveillance aircraft and other sources.

‘Radical acceleration’ in AI military targeting raises oversight fears

Issued on: 10/03/2026
13:18 min

Speaking with FRANCE 24’s Sharon Gaffney, Elke Schwarz, Professor of Political Theory at Queen Mary University of London, warned of a “radical acceleration” in the speed at which military targets are identified through the use of AI, and how quickly action is taken against them. She said this raises concerns about the lack of human oversight, especially given that AI models have “25 to 50% reliability, which means they are wrong very often”.

BY: Sharon GAFFNEY

EU backs ban on AI-generated sexualised deepfakes after Grok backlash

EU states on Friday supported new rules banning AI tools that generate sexualised deepfakes, following a backlash over images made with Elon Musk’s AI chatbot Grok.


Issued on: 13/03/2026 
By: FRANCE 24

The EU move comes after an outcry over sexualised deepfake images of women and minors created by AI chatbot Grok © Pablo VERA / AFP/File

EU nations on Friday backed a ban on AI systems generating sexualised deepfakes, after an outcry over such images produced by Elon Musk's AI chatbot Grok.

European ambassadors agreed to prohibit "practices regarding the generation of non-consensual sexual and intimate content or child sexual abuse material", a spokesperson for Cyprus, which holds the EU's rotating presidency, said.

EU member states and the European Parliament introduced the ban as part of the proposals to amend the bloc's comprehensive rules on AI.

EU lawmakers are set to approve the ban during a vote by committees Wednesday.

TRUTH OR FAKE © FRANCE 24

03:39

"It's not just about individual scandals like Grok. It's about how much power we are willing to give AI to degrade people," EU lawmaker Sergey Lagodinsky, one of several members of the parliament pushing for the ban, said.

X, the platform on which Grok is available, in January said it had "zero tolerance" for sexualised deepfakes of children and women, and implemented measures it said would stop such image creation after the global outrage.

The European Commission, the bloc's digital watchdog, in January kickstarted an investigation into Grok under the EU's online content rules.

The ban will become law after negotiations on a final text including the changes to the AI rulebook between the EU parliament and member states.

Ambassadors on Friday also approved a fixed timeline for the delayed application of high-risk AI rules: December 2027 for stand-alone high-risk AI systems and August 2028 for high-risk AI systems embedded in products.

Rules on high-risk AI are due to come into effect in August 2026 and August 2027 unless the delay becomes law.

(FRANCE 24 with AFP)

Iran trolls Trump with AI-generated LEGO propaganda video

TRUTH OR FAKE © FRANCE 24
07:09




Issued on: 12/03/2026 - FRANCE24

BY: Vedika BAHL

07:09 min

The war in Iran is not only being fought on the battlefield – it is also playing out online. Iran’s state media recently took a leaf out of the White House’s own social media playbook, mocking Donald Trump and Benjamin Netanyahu with an AI-generated propaganda video styled like a LEGO animation. The clip suggested that Trump launched the conflict to distract from scrutiny over his links to disgraced financier Jeffrey Epstein. The video quickly circulated online, highlighting how artificial intelligence is being used as a tool of political messaging and satire in modern conflicts.


Tehran’s video appears to be a direct response to the White House’s own aggressive digital strategy, which uses AI and memes to attack opponents. In recent weeks, Washington’s official accounts have pumped out a stream of viral content about US military action in Iran – splicing real missile-strike footage with memes, pop-culture references and video-game imagery – in an effort to win the narrative battle online and flex its technological and military might.

As governments increasingly turn to shareable content to influence public opinion, distinguishing fact from manipulation becomes more challenging.

In this edition of Truth or Fake, Vedika Bahl analyses how information warfare is unfolding across social platforms and examines the line between messaging, misinformation and digital propaganda in the Middle East war.