Imagen

Training compute

1.5×10²² FLOP

Parameters

7.8B

Published

May 23, 2022

Imagen is an AI model developed by Google Brain (United States), first published in May 2022. It works in the image generation domain, on tasks such as text-to-image and image generation.

Training it took an estimated 1.5×10²² FLOP of compute (estimation method: hardware). The model has 7,762,000,000 parameters. Training ran on 256 Google TPU v4 for about 96 hours. The compute alone is estimated at $8K in 2023 dollars.

Access: API access. Its weights are not openly released. The reference paper has 8,259 citations. Epoch AI rates the confidence of this record as likely.

Full record

Organization: Google Brain
Country of organization: United States
Domain: Image generation
Task: Text-to-image, Image generation
Training compute: 1.5×10²² FLOP
Compute estimation method: Hardware
Parameters: 7,762,000,000
Training hardware: Google TPU v4
Chips used: 256
Training time: 96 h
Chip-hours: 24.6K
Training power draw: 174.7 kW
Training cost (2023 USD): $8K
Model accessibility: API access
Open weights: No
Citations: 8,259
Epoch confidence: Likely