BigGAN-deep 512x512
BigGAN-deep 512x512 is an AI model developed by Heriot-Watt University and DeepMind (United Kingdom), first published in September 2018. It works in the image generation domain, on tasks such as image generation. It counts among the frontier models: the systems trained with the most compute of their moment.
Training it took an estimated 1.8×10²¹ FLOP of compute (estimation method: third-party estimation). The model has 112,694,781 parameters. It was trained on roughly 584M datapoints. Training ran on 256 Google TPU v3 for about 48 hours. The compute alone is estimated at $5K in 2023 dollars.
Access: Open weights (unrestricted). Its weights are openly available. The reference paper has 6,101 citations. Epoch AI rates the confidence of this record as likely.
- Organization
- Heriot-Watt University, DeepMind
- Country of organization
- United Kingdom
- Domain
- Image generation
- Task
- Image generation
- Training compute
- 1.8×10²¹ FLOP
- Compute estimation method
- Third-party estimation
- Parameters
- 112,694,781
- Dataset size
- 584M
- Training hardware
- Google TPU v3
- Chips used
- 256
- Training time
- 48 h
- Chip-hours
- 12.3K
- Training power draw
- 238.2 kW
- Training cost (2023 USD)
- $5K
- Numerical format
- FP32
- Model accessibility
- Open weights (unrestricted)
- Open weights
- Yes
- Citations
- 6,101
- Epoch confidence
- Likely