clip-retrieval
Compute clip embeddings and build a scalable semantic search system for images and text.
Easily compute clip embeddings and build efficient clip retrieval indices.
Clip-retrieval lets you compute multimodal embeddings and build semantic search systems for large image and text datasets. It suits developers and technical professionals organizing massive visual media libraries using AI.
What it is
The tool provides python clients, inference engines to process thousands of samples per second, index builders, and a Flask backend with a web UI. You can query indices by text, image, embedding, or directory of files, and integrate with img2dataset to extract new datasets.
Who it's for
It suits machine learning engineers and developers comfortable with Python, Jupyter notebooks, and command-line interfaces who need to search large collections of images using natural language. It does not suit non-technical users looking for an out-of-the-box desktop video editing application.
Key features
Clip inference
Quickly compute image and text embeddings at rates over 1000 samples per second on consumer GPUs.
Clip client
Query a remote backend via python using text, images, or raw embeddings.
End-to-end pipeline
Run dataset downloading, embedding inference, index building, backend, and UI with simple commands.
Aesthetic scoring
Filter and weight search results using integrated aesthetic predictors and safety models.
Up and running in 3 steps
Our honest verdict
What's great
- Processes millions of multimodal samples rapidly on consumer hardware like a 3080
- Includes a ready-to-use web UI and Flask backend for visual querying
- Supports both single-machine scripts and large-scale distributed index setups
- Integrates directly with img2dataset for building datasets from search outputs
Where it falls short
- Requires familiarity with Python, embeddings, and command-line tools
- Running a large-scale index backend demands dedicated server resources and storage
- GPU required for practical inference speeds on large datasets
Pricing & license
Free under its MIT licence. There are no paid hosted options mentioned in the README, so the only cost is running the hardware and servers yourself.
Open remote AI-training roles
873 open remote AI-training roles, updated November 15, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Can I search using languages other than English?
Yes, by using the multi-lingual version of CLIP (mclip) during client initialization.
How do I handle datasets without a GPU?
You can export CUDA_VISIBLE_DEVICES= to run on CPU, though processing speeds will be significantly lower.
What kind of input formats are supported for inference?
You can use local files or webdataset formats.
