openclip-gemma brings EmbeddingGemma 2 to OpenCLIP-style Python

The project grew out of a personal frustration: the developer uses CLIP regularly but struggled to tune it for specific scene detection. When EmbeddingGemma 2 appeared, it offered image-text comparison like CLIP but was built on different principles, so it could not be used the same way. The wrapper is an attempt to bridge that gap while keeping the official preprocessing, tokenizer, prompt, pooling and normalization logic intact, with model weights pulled from Hugging Face at runtime rather than bundled.

The developer reports close alignment with the official inference path: on Windows 11, Python 3.12 and CPU FP32, the maximum absolute error was about 2.74 × 10⁻⁷ for image embeddings, 2.81 × 10⁻⁶ for text embeddings, and 1.19 × 10⁻⁶ for image-text similarity. Those figures compare against Sentence Transformers inference for a specific model revision and test environment, so they indicate fidelity in that setup rather than proven accuracy on specific classification or retrieval tasks. The pre-release checks covered 75 passing pytest tests with 9 skipped, a real-model CPU integration test, and automated CI and PyPI publishing via GitHub Actions.

The library is positioned narrowly: v0.1.0 supports only EmbeddingGemma 2, handles images and text (not video or audio), and omits some OpenCLIP features such as forward() and get_logits(). The developer plans to widen platform support and improve documentation based on usage, and notes the name may change to respect existing project trademarks, with migration guidance to follow on GitHub.