Google Cloud adds AI labels to stored images and PDFs in new preview
Object enrichments handles model processing and metadata storage inside Cloud Storage. Google quotes $700 for 10 million images or PDFs; Gemini-based tasks remain upcoming.
Loading page…
Object enrichments handles model processing and metadata storage inside Cloud Storage. Google quotes $700 for 10 million images or PDFs; Gemini-based tasks remain upcoming.
Listen to this story
Google Cloud’s object enrichments preview turns image and PDF labeling into a managed Cloud Storage workflow: teams select files and models, and the service processes them and saves results with each object. That can reduce repeated model calls and the custom queues, workers, and metadata stores teams otherwise maintain. The preview is limited to five image and PDF models; Gemini-based summarization, extraction, transcription, and translation are planned separately and require an access request.
Google quotes $700 to enrich 10 million images or PDFs, according to its published pricing.
Object context travels with a file when it is copied, moved, or replicated, and is deleted with the file.
L’Oréal reports labeling creative assets 100 times faster; Hyperscience reports over 50% higher classification accuracy and two- to threefold lower inference costs.
Google Cloud is giving companies a way to prepare stored files for AI agents without building a separate processing pipeline. Its new object enrichments preview extracts descriptive metadata from images and PDFs, then keeps that information alongside the files. The service is available through the Cloud Storage console and API.
The starting problem is repeated work. Google describes an operations agent looking for a package delivery or road hazard: feeding thousands of camera frames into a model for every request consumes tokens and adds delay. Extracting useful descriptions beforehand gives the agent information it can consult at runtime instead. Until now, Google says, teams building that preparation layer have had to coordinate triggers, queues, processing infrastructure and error handling, often with a separate database for the results.
The preview offers five first-party models for images and PDFs. Image tasks include detecting objects and faces, assigning labels and flagging explicit content. The PDF classifier sorts documents into standard business and industry categories, including Finance, Games, News and Health. Google says the image models include technology used by its internal services, such as Google Photos. These are annotation and classification tools, distinct from the broader Gemini capabilities on the roadmap.
Google quotes this cost according to its published pricing.
Customers can configure a job without code in the console or use the Cloud Storage API. Setup has three parts:
After that, the service coordinates model processing, parses the responses and writes results back to each file’s metadata. Retries and throttling are built in. Google says jobs can cover millions or billions of objects, removing the need for customers to manage their own queues or scale processing workers.
The results go into object context, Cloud Storage’s structure for descriptive key-value pairs attached to a stored file. That context follows the object when it is copied, moved or replicated, and disappears when the object is deleted. The design addresses a maintenance problem Google identifies with separate metadata stores: teams otherwise need cleanup jobs to keep descriptions synchronized with the files they describe.
Administrators can use identity and access management permissions to control who reads, adds, changes or removes that context. Applications can filter file listings using its key-value pairs. Customers can also configure the metadata to flow into BigQuery through Storage Insights datasets, allowing analysts to query file-derived information and combine it with structured business tables without writing a separate data-transfer workflow.
Google’s announcement includes early customer claims, rather than an independent performance comparison. L’Oréal principal data and AI architect Antoine Castex says the service lets the company label its creative asset repository 100 times faster than existing processes. Hyperscience’s vice president of partnerships, Mayur Pillay, reports a greater than 50% lift in document-classification accuracy and classification inference costs reduced by a factor of two to three.
Google describes Gemini-based enrichments as upcoming functionality, not part of the currently available five-model offering. Its proposed tasks include summarization, extracting named entities, transcription and translation across text, PDFs, images, audio and video. Customers must contact Google to learn more and request access. For teams starting now, the concrete option is the image and PDF preview; richer multimodal processing remains a separate access request.
Loading discussion...
Join the conversation
Explain when convenience should give way to checking the labels.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.