clip

Generate normalized image and text embeddings for zero-shot classification and retrieval workflows.

Updated Apr 9, 2026

Search Tags:#vector database #image search #content moderation #clip #zero-shot classification #cross-modal retrieval #image-text embeddings

Ready to run

npx skills add https://github.com/MarbleSodas/Mavis --skill clip-marblesodas

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: clip
Source: https://github.com/MarbleSodas/Mavis/tree/main/optional-skills/mlops/clip
Command: npx skills add https://github.com/MarbleSodas/Mavis --skill clip-marblesodas

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill solves the problem of matching images to natural-language descriptions without training custom models.

Core Features & Use Cases

Zero-shot image classification: Assign labels to images using only text prompts, no fine-tuning required.
Cross-modal retrieval: Search for the most relevant images given a text query, and compare image↔text similarity.
Embedding-based workflows: Generate image and text features that can be stored in vector databases for fast retrieval and deduplication use cases.

Quick Start

Run CLIP with your chosen vision-language model (such as ViT-B/32) to compute image and text embeddings, then rank images by cosine similarity to your query.

Frequently Asked Questions about clip

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I classify images using text labels without collecting training data?▼

Zero-shot image classification assigns human-readable text labels to images without training data by aligning image and text embeddings using vision and text encoders, then matching images to your text prompts via cosine similarity.

Can I use CLIP for cross-modal retrieval to find images from text queries?▼

Cross-modal retrieval supports searching relevant images given a text query by computing image and text embeddings, enabling image-to-text and text-to-image similarity matching without requiring labeled training data.

How do I generate image and text embeddings for a vector database?▼

Generate image and text embeddings for a vector database by running a vision-language model like ViT-B/32 through CLIP, applying embedding normalization to produce deterministic features compatible with fast retrieval indexes.

What's the best way to perform image deduplication and content filtering?▼

Image deduplication and content filtering use embedding-based workflows by extracting deterministic image features with vision encoders, then comparing normalized embeddings stored in a vector database to identify duplicates.

Do I need a fine-tuned model to match images to natural-language descriptions?▼

No fine-tuned model is required to match images to natural-language descriptions because CLIP aligns images with human-readable labels directly, enabling zero-shot classification and image-text matching out of the box.

Can I use CLIP for content moderation without labeled training data?▼

Content moderation without labeled training data is possible using zero-shot classification by defining text prompts for moderation categories, then ranking images by cosine similarity to those text embeddings.