SILVA Aesthetic Scorer

SILVA Aesthetic Scorer

Scores an uploaded illustration three ways, with one person's taste and two distilled AI judges, each as a number from 0 to 1. Suits artists, dataset curators and researchers of taste models.

Made by Jannchie on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This Space scores an illustration for how appealing it looks. You upload a picture and get three scores between 0 and 1. One reflects a single person's taste, and the other two come from AI judges whose opinions were distilled into small heads. All three read the same image embedding, so the heavy part runs once.

Who it suits: It suits illustrators who are curious how a model rates their work, and people who sort large image sets and want a taste signal to rank them.

Industry | Who uses it | How they use it | Example
Illustration | Artists | See how a model rates a finished picture | Upload a drawing and read its three scores
Data curation | Machine learning engineers | Rank a large image set by taste | Score a folder of pictures with the command line tool
Research | Researchers | Compare how different judges agree and disagree | Compare the person's score with the two AI scores
Communities | Gallery curators | Add a taste signal when shortlisting pictures | Score a shortlist and look at the order

Skip it if you want an objective measure of quality, since the page says each head reflects one consistent taste.

Where to use it:

• A browser: open the Space's page on Hugging Face, with nothing to install
• A copy on Windows, macOS or Linux, using Python
• A terminal, through the silva-scorer package and its command line tool

Use case | Who | What they get back
Scoring one illustration | Artists | Three scores from 0 to 1 for the same picture
Ranking a set | Curators | A score for each image you run through the tool
Comparing judges | Researchers | Side by side scores from a person-trained head and two judge heads

How to set it up: The online copy needs nothing. For a copy on your own computer, you need the following.

• Python with pip on Windows, macOS or Linux
• The Hugging Face hub client, which the first step installs
• The packages listed in requirements.txt

1. Install the Hugging Face hub client.
2. Download the Space's files into a local folder.
3. Move into that folder.
4. Install Gradio.
5. Install the remaining packages from requirements.txt.
6. Start app.py.

pip install -U huggingface_hub
hf download Jannchie/silva-aesthetic-demo --repo-type space --local-dir silva-aesthetic-demo
cd silva-aesthetic-demo
pip install gradio
pip install -r requirements.txt
python app.py

To score images from a terminal, install the package and run its tool on a file.

pip install "silva-scorer[backbone]"
silva score image.jpg --repo-id Jannchie/silva-aesthetic

How to use it: In the Space, upload an illustration and read the three scores. The page says the first run after a cold start is slow because the backbone must load, which costs several tens of seconds on its free tier. Later runs are quick. In a terminal, use the tool as shown above.

Tool | What it does
silva score | Scores an image file with the head from the repository you name
Space page | Shows three scores for one uploaded illustration

Example: You upload a character illustration to the Space. After a short wait it shows three numbers between 0 and 1, one for the person's taste and two from the AI judges. The source gives no sample values.

For newcomers and experts: If you are new, upload one picture and wait for the first result, since the first run is the slowest. Then try a second picture and compare how the three scores differ.

If you know the field, a frozen SigLIP2 backbone turns the image into a 1152-dimension embedding, and small ordinal-regression heads of about 7 MB map it to calibrated scores. Each head is calibrated to one rater's 1 to 5 label distribution. Across 229,129 pictures the two judges agree closely with each other but only loosely with the person.

Title: Spearman agreement the page reports
Judge to judge: 0.825
Judge to person, lower end: 0.45
Judge to person, upper end: 0.48

You upload an illustration
The frozen backbone makes one embedding
Three small heads each read that embedding
You see three scores from 0 to 1

Good to know: Read each score as one consistent taste, not a universal measure of quality.

• License: mit, as listed on Hugging Face
• Cost to run: no API key or paid service is named, and the online copy runs on cpu-basic hardware
• Data it reaches: the pictures you upload, which the online Space handles on Hugging Face's servers, and a local copy downloads the Space's files
• Not for: the source lists no exclusions, but the scores are not a universal measure of picture quality

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: gradio
• License: mit

Links

• Space: https://huggingface.co/spaces/Jannchie/silva-aesthetic-demo
• Files: https://huggingface.co/spaces/Jannchie/silva-aesthetic-demo/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.