pizza-clip-base-v1

pizza-clip-base-v1 is a production-oriented ONNX export of the CLIP ViT-B/32 vision and text encoders, published by Infinilabs for use with Pizza Engine and other local retrieval pipelines.

It provides aligned 512-dimensional embeddings for:

• image to image retrieval
• text to image retrieval
• image to text matching
• cross-modal search demos in local or server-side inference flows

Model origin

This model family is derived from openai/clip-vit-base-patch32 and repackaged as standalone ONNX artifacts for practical deployment with ONNX Runtime.

Files

• clip-vit-base-patch32-vision.onnx: vision encoder graph
• clip-vit-base-patch32-vision.onnx.data: external vision weights
• clip-vit-base-patch32-text.onnx: text encoder graph
• clip-vit-base-patch32-text.onnx.data: external text weights
• tokenizer.json: tokenizer for text encoding

Embedding shape

• dimension: 512
• image input: pixel_values with shape (batch, 3, 224, 224)
• text inputs: input_ids, attention_mask

Both encoders return L2-normalized embeddings so cosine similarity can be used directly for ranking.

Intended usage

This repo is intended for inference, search, and retrieval workloads, especially:

• local ONNX Runtime inference
• Rust-based retrieval pipelines
• multimodal search prototypes
• lightweight deployment without Python model loading in production

Example use cases

• retrieve poster-like images from a text query such as historical drama poster
• rank captions against cover art
• build cross-modal search in Pizza Engine

Notes

• This repository contains exported inference artifacts, not training code.
• Tokenization behavior follows the upstream CLIP tokenizer.
• Accuracy and behavior should track the upstream CLIP ViT-B/32 model within normal ONNX export tolerance.

Integration

Pizza Engine can use this model family for cross-modal demos alongside text embeddings such as pizza-embed-zh-small-v1.

Attribution

• upstream model: openai/clip-vit-base-patch32
• packaging and ONNX deployment artifacts: Infinilabs