Jul 2025 – Oct 2025

Jul 2025 – Oct 2025

• Product Design Intern

• Product Design Intern

Interned with Shanghai Lab’s Frontier Exploration Center, collaborating with Alibaba Cloud and Huawei to turn AI research into user-facing tools. Led 3 projects, including ArtiMuse, an AI aesthetics tool launched at WAIC 2025 with 20K+ early users. Online: https://artimuse.intern-ai.org.cn

Interned with Shanghai Lab’s Frontier Exploration Center, collaborating with Alibaba Cloud and Huawei to turn AI research into user-facing tools. Led 3 projects, including ArtiMuse, an AI aesthetics tool launched at WAIC 2025 with 20K+ early users. Online: https://artimuse.intern-ai.org.cn

Interned with Shanghai Lab’s Frontier Exploration Center, collaborating with Alibaba Cloud and Huawei to turn AI research into user-facing tools. Led 3 projects, including ArtiMuse, an AI aesthetics tool launched at WAIC 2025 with 20K+ early users. Online: https://artimuse.intern-ai.org.cn

Team

Team

Team

Minzhe Emily

Minzhe Emily

Minzhe Emily

Product design intern

Product design intern

Product design intern

Product manager

Product manager

Product manager

Cao Li

Cao Li

Cao Li

Development engineer

Development engineer

Development engineer

Fubin

Fubin

Fubin

Mentor

Yihao Liu

Yihao Liu

Yihao Liu

Team leader

Team leader

Team leader

Skills

Skills

Skills

Product thinking & strategy

Product thinking & strategy

Product thinking & strategy

User research

User research

User research

Brainstorming

Brainstorming

Brainstorming

UI / UX Design

UI / UX Design

UI / UX Design

Prototyping

Prototyping

Prototyping

Timeline

Figma

Lark

Tools

4 Weeks

Timeline

Timeline

4 Weeks

4 Weeks

Tools

Tools

Figma

Figma

Lark

Lark

Background
Background
Background

Why we built ArtiMuse

Why we built ArtiMuse

Why we built ArtiMuse

We found that many existing AI models, such as ChatGPT, MidJourney, and Stable Diffusion, are impressive at generating content, but they still fall short in evaluating aesthetics. They often provide only vague scores or unprofessional comments, lacking clear standards and detailed analysis. This means that people’s needs remain unmet, and a significant white space still exists in the market. In fact, the demand for aesthetic evaluation is widely present across multiple domains.

We found that many existing AI models, such as ChatGPT, MidJourney, and Stable Diffusion, are impressive at generating content, but they still fall short in evaluating aesthetics. They often provide only vague scores or unprofessional comments, lacking clear standards and detailed analysis. This means that people’s needs remain unmet, and a significant white space still exists in the market. In fact, the demand for aesthetic evaluation is widely present across multiple domains.

We found that many existing AI models, such as ChatGPT, MidJourney, and Stable Diffusion, are impressive at generating content, but they still fall short in evaluating aesthetics. They often provide only vague scores or unprofessional comments, lacking clear standards and detailed analysis. This means that people’s needs remain unmet, and a significant white space still exists in the market. In fact, the demand for aesthetic evaluation is widely present across multiple domains.

Familiar challenges we all face

Familiar challenges we all face

Familiar challenges we all face

Your boyfriend took unflattering pictures

Your boyfriend took
unflattering pictures

Your boyfriend took unflattering pictures

ArtiMuse helps your partner with aesthetic insights to take better photos!

ArtiMuse helps your partner with aesthetic insights to take better photos!

ArtiMuse helps your partner with aesthetic insights to take better photos!

1

As a designer, clients picked your first drafts after endless edits

As a designer, clients picked your first drafts after endless edits

As a designer, clients picked your first drafts after endless edits

ArtiMuse offers aesthetic guidance to lock onto the right direction and cutting down endless revisions!

ArtiMuse offers aesthetic guidance to lock onto the right direction and cutting down endless revisions!

ArtiMuse offers aesthetic guidance to lock onto the right direction and cutting down endless revisions!

2

As a drawing learner, finding real guidance is hard & expensive

As a drawing learner, finding real guidance is hard & expensive

As a drawing learner, finding real guidance is hard & expensive

ArtiMuse offers instant expert aesthetic feedback to sharpen your drawing skills!

ArtiMuse offers instant expert aesthetic feedback to sharpen your drawing skills!

ArtiMuse offers instant expert aesthetic feedback to sharpen your drawing skills!

3

Looking at an artwork, you want
to know its craft and meaning

Looking at an artwork, you want to
know its craft and meaning

Looking at an artwork, you want to know its craft and meaning

ArtiMuse makes art appreciation easier by offering instant insights into each piece’s aesthetic value!

ArtiMuse makes art appreciation easier by offering instant insights into each piece’s aesthetic value!

ArtiMuse makes art appreciation easier by offering instant insights into each piece’s aesthetic value!

4

From pain points to potential

From pain points to potential

From pain points to potential

From these challenges, we partnered with the China Academy of Art to create ArtiMuse, an expert-level aesthetic AI model that provides trustworthy scores and feedback, helping users in learning, creation, and daily expression.

From these challenges, we partnered with the China Academy of Art to create ArtiMuse, an expert-level aesthetic AI model that provides trustworthy scores and feedback, helping users in learning, creation, and daily expression.

From these challenges, we partnered with the China Academy of Art to create ArtiMuse, an expert-level aesthetic AI model that provides trustworthy scores and feedback, helping users in learning, creation, and daily expression.

SCREENSHOTS

Pepper Screen
Pepper Screen
Pepper Screen
Pepper Screen
Pepper Screen
Pepper Screen
Pepper Screen
Pepper Screen
Pepper Screen
Predict Screen
Predict Screen
Predict Screen

AI x ART: Expert-level aesthetic model

AI x ART: Expert-level aesthetic model

AI x ART: Expert-level aesthetic model

Advanced AI foundation

Shanghai AI Lab powers ArtiMuse’s core technology with multimodal AI and large-scale training.

Academic collaboration

Partnering with top art universities, professors and researchers built a large-scale aesthetic dataset through detailed annotation.

Academic collaboration

Partnering with top art universities, professors and researchers built a large-scale aesthetic dataset through detailed annotation.

Leading the field

Outperforming

leading models like

GPT-4 and InternVL.

Advanced AI foundation

Advanced AI foundation

Shanghai AI Lab powers ArtiMuse’s core technology with multimodal AI and large-scale training.

Shanghai AI Lab powers ArtiMuse’s core technology with multimodal AI and large-scale training.

Academic collaboration

Partnering with top art universities, professors and researchers built a large-scale aesthetic dataset through detailed annotation.

Academic collaboration

Partnering with top art universities, professors and researchers built a large-scale aesthetic dataset through detailed annotation.

Leading the field

Outperforming

leading models like

GPT-4 and InternVL.

Outperforming

leading models like

GPT-4 and InternVL.

Composition

Composition

Technique

Technique

Visuals

Visuals

Overall evaluation

Overall evaluation

Creativity

Creativity

Emotion

Emotion

Form

Form

Theme

Theme

AI evaluation logic

AI evaluation logic

AI evaluates design quality through 8 aesthetic dimensions, revealing the reasons behind each score.

Composition

Composition

Technique

Technique

Visuals

Visuals

Overall evaluation

Overall evaluation

Creativity

Creativity

Emotion

Emotion

Form

Form

Theme

Theme

AI evaluation logic

AI evaluates design quality through 8 aesthetic dimensions, revealing the reasons behind each score.

AI aesthetic evaluation performs

Score: 69 / 100

The composition is dynamic, with fruits and draped fabric creating movement. Contrasting colors add visual interest. Though the background slightly competes with the subject, the overall layout remains harmonious and cohesive.

blue spiral light on red background
a close up of a cell phone with a blurry background

Score: 69 / 100

The composition is dynamic, with fruits and draped fabric creating movement. Contrasting colors add visual interest. Though the background slightly competes with the subject, the overall layout remains harmonious and cohesive.

blue spiral light on red background
a close up of a cell phone with a blurry background

AI aesthetic evaluation performs

Aesthetics score: 69 /100

The composition is dynamic and balanced, with fruits and draped fabric creating movement. Contrasting colors and textures add visual interest, while the central bowl anchors the scene. Though the background slightly competes with the subject, the overall layout remains harmonious and cohesive.

a close up of a cell phone with a blurry background
blue spiral light on red background

3 Steps to discover your aesthetic DNA

3 Steps to discover your

aesthetic DNA

AI aesthetic evaluation performs

Score: 69 / 100

The composition is dynamic, with fruits and draped fabric creating movement. Contrasting colors add visual interest. Though the background slightly competes with the subject, the overall layout remains harmonious and cohesive.

blue spiral light on red background
a close up of a cell phone with a blurry background

Score: 69 / 100

The composition is dynamic, with fruits and draped fabric creating movement. Contrasting colors add visual interest. Though the background slightly competes with the subject, the overall layout remains harmonious and cohesive.

blue spiral light on red background
a close up of a cell phone with a blurry background

3 Steps to discover your aesthetic DNA

3 Steps to discover your

aesthetic DNA

This flow is a sample scenario for a bank’s app, but the FaceID verification feature can be seamlessly integrated into a wide range of products.

Who’s scoring you today?

Who’s scoring you today?

Your judges might look familiar, perhaps Picasso, Monet, or another legendary master. ArtiMuse generates art characters that mirror your work’s style, holding a score card and making every evaluation a moment of playful surprise.

Your judges might look familiar, perhaps Picasso, Monet, or another legendary master. ArtiMuse generates art characters that mirror your work’s style, holding a score card and making every evaluation a moment of playful surprise.

  • shoot 2
  • shoot 1
  • shoot 4
  • shoot 3
  • shoot 1
  • shoot 2
  • shoot 3
  • shoot 4

Who’s scoring you today?

Your judges might look familiar, perhaps Picasso, Monet, or another legendary master. ArtiMuse generates art characters that mirror your work’s style, holding a score card and making every evaluation a moment of playful surprise.

  • shoot 2
  • shoot 1
  • shoot 4
  • shoot 3
  • shoot 1
  • shoot 2
  • shoot 3
  • shoot 4

Meet the sassy mode

I realized that one of the biggest reasons people don’t fully trust AI evaluation is because it often feels…fake. AI loves to flatter you “Great job! Amazing work!” But after hearing it a dozen times, it just sounds like empty words. ArtiMuse takes a different approach. A sassy mode you can switch to. Where instead of endless praise, you actually get sharp, sassy, sometimes even a little mean feedback😉.

What people seek from AI isn’t praise, but to be understood and felt.

Okay, but I’ve seen hotter.🔥🔥🔥

Okay, but I’ve seen hotter.🔥🔥🔥

You clearly closed the book right before the grid system chapter.

You clearly closed the book right before the grid system chapter.

Oh, DALL·E taught you that? Adorable.

Oh, DALL·E taught you that? Adorable.

This felt like you were playing it safe. Try pushing boundaries next time.

This felt like you were playing it safe. Try pushing boundaries next time.

+0%

+0%

By the numbers

87% of users said Sassy Mode made feedback feel fresher and more charming, with positivity up 33.6%.

+0%

+0%

By the numbers

87% of users said Sassy Mode made feedback feel fresher and more charming, with positivity up 33.6%.

Polite? Sassy? You decide.

Solving traditional AI model pain points

1. More transparent evaluation

Previous models generated scores without explaining why. Even text feedback was templated and lacked meaningful insights.

1. More transparent evaluation

Previous models generated scores without explaining why. Even text feedback was templated and lacked meaningful insights.

2. Reducing cultural bias

Previous models were largely trained on Western art and failed to interpret artistic logic from Asia, Africa, and other regions.

2. Reducing cultural bias

Previous models were largely trained on Western art and failed to interpret artistic logic from Asia, Africa, and other regions.

3. Deeper insight

Previous models stayed at the surface level like “good looking”, often favoring flattering content and producing inconsistent scores.

3. Deeper insight

Previous models stayed at the surface level like “good looking”, often favoring flattering content and producing inconsistent scores.

Meet the sassy mode

Meet the sassy mode

I realized that one of the biggest reasons people don’t fully trust AI evaluation is because it often feels…fake. AI loves to flatter you “Great job! Amazing work!” But after hearing it a dozen times, it just sounds like empty words. ArtiMuse takes a different approach. A sassy mode you can switch to. Where instead of endless praise, you actually get sharp, sassy, sometimes even a little mean feedback😉.

I realized that one of the biggest reasons people don’t fully trust AI evaluation is because it often feels…fake. AI loves to flatter you “Great job! Amazing work!” But after hearing it a dozen times, it just sounds like empty words. ArtiMuse takes a different approach. A sassy mode you can switch to. Where instead of endless praise, you actually get sharp, sassy, sometimes even a little mean feedback😉.

What people seek from AI isn’t praise, but to be understood and felt.

What people seek from AI isn’t praise, but to be understood and felt.

Okay, but I’ve seen hotter.🔥🔥🔥

Okay, but I’ve seen hotter.🔥🔥🔥

You clearly closed the book right before the grid system chapter.

You clearly closed the book right before the grid system chapter.

Oh, DALL·E taught you that? Adorable.

Oh, DALL·E taught you that? Adorable.

This felt like you were playing it safe. Try pushing boundaries next time.

This felt like you were playing it safe. Try pushing boundaries next time.

+0%

+0%

By the numbers

87% of users said sassy mode made the feedback feel fresher and charm, the positive sentiment increased by 33.6%.

+0%

+0%

By the numbers

87% of users said sassy mode made the feedback feel fresher and charm, the positive sentiment increased by 33.6%.

Polite? Sassy? You decide.

Polite? Sassy? You decide.

Solving traditional model pain points

Solving traditional model pain points

More transparent evaluation

Previous models generated scores without explaining why. Even text feedback was templated and lacked meaningful insights.

Deeper insight

Previous models stayed at the surface level like “good looking”, often favoring flattering content and producing inconsistent scores.

Reducing cultural bias

Previous models were largely trained on Western art and failed to interpret artistic logic from Asia, Africa, and other regions.

Deeper insight

Previous models stayed at the surface level like “good looking”, often favoring flattering content and producing inconsistent scores.

More transparent evaluation

Previous models generated scores without explaining why. Even text feedback was templated and lacked meaningful insights.

Deeper insight

Previous models stayed at the surface level like “good looking”, often favoring flattering content and producing inconsistent scores.

Reducing cultural bias

Previous models were largely trained on Western art and failed to interpret artistic logic from Asia, Africa, and other regions.

Deeper insight

Previous models stayed at the surface level like “good looking”, often favoring flattering content and producing inconsistent scores.

Application scenarios

Application scenarios

Application scenarios

Across diverse art forms

AIGC

Painting

Photography

Calligraphy

Graphic design

3D

Across diverse art forms

Across diverse art forms

AIGC

Painting

Photography

Calligraphy

Graphic design

3D

AIGC

Painting

Photography

Calligraphy

Graphic design

3D

Loved by 12K+ users upon release

Loved by 12K+ users upon release

Loved by 12K+ users upon release

Released at WAIC 2025 with an expert-curated 10K dataset, ArtiMuse attracted 12K+ early users and received highly positive feedback. It surpassed previous models across the board in fine-grained evaluation.

Released at WAIC 2025 with an expert-curated 10K dataset, ArtiMuse attracted 12K+ early users and received highly positive feedback. It surpassed previous models across the board in fine-grained evaluation.

Released at WAIC 2025 with an expert-curated 10K dataset, ArtiMuse attracted 12K+ early users and received highly positive feedback. It surpassed previous models across the board in fine-grained evaluation.

Behind the scene process

Bridging AI & humans

Bridging AI & humans

While building ArtiMuse’s 8-dimensional framework, I tackled 2 challenges: helping AI learn from experts and helping users understand AI. I worked with art schools and AI teams to translate expert insights into machine-learnable logic and visualize the outcomes for clearer evaluation.

Identifying users’ pain points in AI

Identifying users’ pain points in AI aesthetics

I analyzed user feedback on existing AI aesthetic tools and identified 3 recurring pain points: overly technical language, empty flattering tone, and western-centric bias.

Refine model training

Science-driven model training

ArtiMuse is grounded in a rigorous scientific process, combining curated datasets, expert annotations, and multi-stage model training. I participated in the model training phase, comparing AI’s evaluations across artworks to iterate its aesthetic logic and feedback clarity.

Bridging AI & humans

Bridging AI & humans

While building ArtiMuse’s 8-dimensional framework, I tackled 2 challenges: helping AI learn from experts and helping users understand AI. I worked with art schools and AI teams to translate expert insights into machine-learnable logic and visualize the outcomes for clearer evaluation.

Identifying users’ pain points in AI

Identifying users’ pain points in AI aesthetics

I analyzed user feedback on existing AI aesthetic tools and identified 3 recurring pain points: overly technical language, empty flattering tone, and western-centric bias.

Refine model training

Science-driven model training

ArtiMuse is grounded in a rigorous scientific process, combining curated datasets, expert annotations, and multi-stage model training. I participated in the model training phase, comparing AI’s evaluations across artworks to iterate its aesthetic logic and feedback clarity.

Abstract

What I learned

My ArtiMuse journey reminded me how science and design enrich one another. Every partner from AI, art, and research worked with rigor and passion to keep the project moving forward. I especially thank my mentor, Yihao Liu, for his constant guidance and support, as well as all the collaborators for their trust and amazing teamwork. Here are my main takeaways:

Bridging science & design

I learned how to let design participate in the construction of data structures, annotation processes, and model logic.

Interdisciplinary integration

Collaborating with art professors, engineers, product researchers taught me to quickly identify core issues across disciplines and turn diverse insights to design strategies.

Humanizing data

Translating algorithmic outputs into perceivable visual and linguistic forms through a design perspective.

Next projects

TikTok Shop · Internship

Nov 2024 - May 2025

TikTok Shop · Internship

Nov 2024 - May 2025

Synchroni neurotech · Startup

2024 - Present

Synchroni neurotech · Startup

2024 - Present

Behind the scene process

Behind the scene process

Bridging AI & humans

Bridging AI & humans

While building ArtiMuse’s 8-dimensional framework, I tackled 2 challenges: helping AI learn from experts and helping users understand AI. I worked with art schools and AI teams to translate expert insights into machine-learnable logic and visualize the outcomes for clearer evaluation.

Identifying users’ pain points in AI aesthetics

Identifying users’ pain points in AI aesthetics

I analyzed user feedback on existing AI aesthetic tools and identified 3 recurring pain points: overly technical language, empty flattering tone, and western-centric bias.

Refine model training

Science-driven model training

ArtiMuse is grounded in a rigorous scientific process, combining curated datasets, expert annotations, and multi-stage model training. I participated in the model training phase, comparing AI’s evaluations across artworks to iterate its aesthetic logic and feedback clarity.

Bridging AI & humans

Bridging AI & humans

While building ArtiMuse’s 8-dimensional framework, I tackled 2 challenges: helping AI learn from experts and helping users understand AI. I worked with art schools and AI teams to translate expert insights into machine-learnable logic and visualize the outcomes for clearer evaluation.

Identifying users’ pain points in AI

Identifying users’ pain points in AI aesthetics

I analyzed user feedback on existing AI aesthetic tools and identified 3 recurring pain points: overly technical language, empty flattering tone, and western-centric bias.

Refine model training

Science-driven model training

ArtiMuse is grounded in a rigorous scientific process, combining curated datasets, expert annotations, and multi-stage model training. I participated in the model training phase, comparing AI’s evaluations across artworks to iterate its aesthetic logic and feedback clarity.

Abstract

What I learned

What I learned

My ArtiMuse journey reminded me how science and design enrich one another. Every partner from AI, art, and research worked with rigor and passion to keep the project moving forward. I especially thank my mentor, Yihao Liu, for his constant guidance and support, as well as all the collaborators for their trust and amazing teamwork. Here are my main takeaways:

My ArtiMuse journey reminded me how science and design enrich one another. Every partner from AI, art, and research worked with rigor and passion to keep the project moving forward. I especially thank my mentor, Yihao Liu, for his constant guidance and support, as well as all the collaborators for their trust and amazing teamwork. Here are my main takeaways:

Bridging science & design
Bridging science & design

I learned how to let design participate in the construction of data structures, annotation processes, and model logic.

Interdisciplinary integration
Interdisciplinary integration

Collaborating with art professors, engineers, product researchers taught me to quickly identify core issues across disciplines and turn diverse insights to design strategies.

Humanizing data
Humanizing data

Translating algorithmic outputs into perceivable visual and linguistic forms through a design perspective.

Next projects

Next projects

TikTok Shop · Internship

Nov 2024 - May 2025

TikTok Shop · Internship

Nov 2024 - May 2025

Synchroni neurotech · Startup

2024 - Present

Synchroni neurotech · Startup

2024 - Present

I'd love to get in touch! Say hi at minzhezou00@gmail.com

I'd love to get in touch! Say hi at minzhezou00@gmail.com

I'd love to get in touch! Say hi at minzhezou00@gmail.com

I'd love to get in touch! Say hi at minzhezou00@gmail.com

Create a free website with Framer, the website builder loved by startups, designers and agencies.