Skip to content

CLIP

ClipEmbed

Computes a CLIP embedding for each image in the dataset using the openai/clip-vit-large-patch14 model.

CLIP embedding is quite fast and well worth it, as it unlocks a lot of functionality.

This node computes CLIP embeddings efficiently in batch, and supports multi-GPU operation. Several other nodes depend on CLIP embeddings, for example FuzzyDedupe and SmartHumanFilter.

You may also use beprepared to compute CLIP embeddings for your own purposes, which can be consumed out of the image.json sidecars in the output directory.

Parameters

  • batch_size (default: 128): The number of images to process in parallel.
  • target_property (default: "clip"): The property to store the CLIP embedding in.

Output properties

  • image.{target_property}: The CLIP embedding for the image.

Example

dataset >> ClipEmbed