Key results
The company
Refik Anadol Studio
refikanadolstudio.comMedia arts and design studio specializing in AI-driven architectural installations.
Result highlights
- Data processing time cut from 30 hours to 80 mins
- 7x faster image embedding generation
- 40% higher caption accuracy vs open-source model
The story
A media art studio preparing a permanent exhibition powered by an AI model trained on one of the world's largest datasets of the natural world.
Processing the massive image archive was prohibitively slow, with test batches taking 30 hours to complete. Previous captioning tools produced only generic tags, lacking the scientific specificity needed to create a knowledgeable AI collaborator.
The studio orchestrated a parallelized pipeline on Vertex AI to handle object detection and generate image embeddings as a single reproducible process. Gemini 2.5 Flash generates scientifically accurate captions, while BigQuery manages metadata to allow instant searching across billions of records. This architecture separates raw file storage from searchable metadata to enable high-speed Retrieval-Augmented Generation workflows.
Quotes
“With Unsupervised, we proved that a machine could dream. Now, with the Large Nature Model, we are teaching it to understand. This transition — from abstract hallucinations to scientific accuracy — is only possible because of the speed and intelligence of this new infrastructure. It allows us to close the gap between human vision and machine capability, creating a partnership that is finally free from limitations.”
Implementation partner
Zencore
zencore.devSupported the studio in building its new data processing technique on Google Cloud.