Imgbeddings: Generate CLIP image embeddings in Python

Understand the project

Practical overview

Use it when a Python workflow needs CLIP-based image vectors and you want an ONNX-based option that does not require PyTorch or TensorFlow.

Python developers preparing image embeddings for downstream machine-learning workflows.
Teams building image clustering, similarity or embedding-search systems.

Generate vectors from individual images or batches of image files and PIL images.
Prepare image vectors for clustering or similarity comparisons.
Feed generated embeddings into a separate vector-search system or downstream classifier.

Before you choose

The package is distributed through PyPI, and its public GitHub repository specifies the MIT license.

The default patch32 model is 88 MB and returns 768-dimensional NumPy vectors.
Patch16 produces more granular embeddings but takes about three times as long as patch32; patch14 is about 3.5 times the model size and three times slower than patch16.
Images are padded and resized to square rather than center-cropped; very wide or tall images may yield less robust embeddings.
The library handles image data, not combined image-and-text embeddings.
The README flags bias in CLIP and advises against production use of the model.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Install imgbeddings, load an image and obtain its CLIP embedding vector.

Install the package with the command below.
Documented command
```
pip3 install imgbeddings
```
Load an image as a PIL image, or provide an image filename.
Import imgbeddings and create an instance with ibed = imgbeddings().
Pass the image to ibed.to_embeddings(image).

A 768-dimensional NumPy embedding vector for the image; the method can also accept lists of filenames or PIL images for batch generation.

The default model is the 88 MB patch32 variant.
Images are padded and resized to square; images with an aspect ratio above roughly 3:1 may produce less robust embeddings.
The project warns about CLIP bias and advises against production use of the model.

Evidence and transparency

Source record details

Source availability: Active
Sources checked: 18 October 2026
Original website title: minimaxir/imgbeddings · Official homepage title · retained in its source language