An assitant to Album Photo Selecting powered by AI
Do you need to select photos for an album, and you have too many photos?
All the process implemented - from upload or remove photos, to splitting to groups by semantic similarity, and to assistance with a chat-bot for better filtering!
You can filter out duplicate photos or similar poses by dividing the photos into clusters of similar photos and poses, and use a chatbot to reduce the number of photos in each cluster to one.
- 8G RAM with GPU or 32G RAM without GPU
- GPU - minimum 8G VRAM
- Intel i7 or Amd Ryzen 7 processors and above
- 16G RAM or more
- Nvidia GeForce RTX 4060 8G-VRAM and above
cd into project directory:
cd /project/folder
Build the project:
docker build -t image_selector .
Run with supported gpu:
docker run -d --gpus all -e USE_GPU=True -e BATCH_SIZE=8 -p 8501:8501 image_selector
Run without supported gpu:
docker run -d -p 8501:8501 image_selector
In browser, open localhost:8501.
Upload photos using Drag&Drop,
or Upload Button.
Wait for processing all of the photos.
The photos will show in Gallery Format divided into clusters, you can delete every photo that you want to filter in your album, above the gallery you can control the clustering, by select between KMeans and DBSCAN algorithms, and control the properties by the Slider.
In the Left Side-Bar you can see the Chat-Bot, it is configured to be Personal Album Assistat, you can select multiple photos by the Check-Box and ask the chat about them, you can ask questions like 'Which photo is better?' 'Is this photo suitable to children album?' and more.
In up of the Left Side-Bar you can find the button to deselecting all the photos that selected by Check-Box to using in the chat-bot.
The Ui implemented by using Streamlit.io framework, this framework desined to Data-Science Visualization projects and it's good for controling in the components.
We chose to using Jina-CLIP V2, that described on this paper - https://arxiv.org/abs/2412.08802.
The benefits of this model is that it trained on pairs of textual description and images that makes it good for retrieval letants with a Semantic attention.
You can read here: https://en.wikipedia.org/wiki/K-means_clustering about the algorithm.
We chose to mesure the distance between the vectors by Cosine Similarity.
You can read here: https://en.wikipedia.org/wiki/DBSCAN about the algorithm.
In this implementation, clusters smaller than a configurable threshold are not removed, unlike the original DBSCAN algorithm.
We chose to mesure the distance between the vectors by Cosine Similarity.
We chose to using QWEN 2.5 VL, that described on this paper - https://arxiv.org/abs/2502.13923.
The benefits of this model is that it very good to aware small details in pictures - even text understanding as describe in the paper (it's happen by using almost all the letents that's coming from the vision embedder), and it much smaller to running locally, in this project for security it implemented to running locally so it is big advantage.
- streamlit_app.py - Streamlit UI code
- Dockerfile - Main Docker File
- requirements.txt - Python dependencies
- src
- chat_impl - Chat implementation
- image_clusterer - Clustering algorithms implementation
- distance.py - DBSCAN implementation
- k_means.py - K-Means implementation
- image_embedder - Embedding image to latent model implementation
- test.ipynb - Jupyter notebook to test clustering algorithms