clip-vit-large-patch14

by openai

Contrastive Language-Image Pre-Training for zero-shot visual understanding

28.6Mdl/month

1,985likes

428Mparams

HuggingFace Use in Pipeline

Identifiers

Model ID

openai/clip-vit-large-patch14

Feature URI

mixpeek://video_descriptor@v1/openai_clip_large_v1

Overview

CLIP (Contrastive Language-Image Pre-Training) is a neural network trained on 400M image-text pairs from the internet. It learns visual concepts from natural language supervision, enabling zero-shot transfer to downstream tasks without task-specific training data.

On Mixpeek, CLIP powers visual embedding extraction, converting video frames and images into 768-dimensional vectors that capture semantic meaning. This enables similarity search across visual content using natural language queries.

Architecture

Vision Transformer (ViT-L/14) with 24 layers, 1024-dim hidden size, 16 attention heads. Text encoder is a 12-layer transformer. Both encoders project into a shared 768-dim embedding space via contrastive learning.

Mixpeek SDK Integration

import { Mixpeek } from "mixpeek";

const mx = new Mixpeek({ apiKey: "API_KEY" });

// Upload and extract visual embeddings
await mx.collections.ingest({
  collection_id: "my-collection",
  source: { url: "https://example.com/video.mp4" },
  feature_extractors: [{
    name: "image_embedding",
    version: "v1",
    params: {
      model_id: "openai/clip-vit-large-patch14"
    }
  }]
});

Capabilities

Zero-shot image classification without fine-tuning
Cross-modal text-to-image and image-to-text retrieval
768-dimensional dense vector embeddings
Processes 224x224 pixel image patches
Supports 40+ languages via multilingual text encoder

Use Cases on Mixpeek

Visual search across video libraries, find frames matching natural language descriptions

Content moderation, detect brand logos, inappropriate content, or specific objects

E-commerce product matching, find visually similar products across catalogs

Media asset management, auto-tag and organize image/video archives

Benchmarks

Dataset	Metric	Score	Source
ImageNet zero-shot	Top-1 Accuracy	75.3%	Radford et al., 2021 — Table 11
MS-COCO (text→image)	Recall@5	56.4%	Radford et al., 2021 — Table 8
Flickr30k (text→image)	Recall@1	87.1%	Radford et al., 2021 — Table 8