Stable Diffusion is a powerful open-source image AI that competes with OpenAI's DALL-E 2. The AI training was probably rather cheap in comparison.
Anyone interested can download the model of the open-source image AI Stable Diffusion for free from Github and run it locally on a compatible graphics card. This must be reasonably powerful (at least 5.1 GB VRAM), but you don't need a high-end computer.
In addition to the local, free version, the Stable Diffusion team also offers access via a web interface. For about $12, you get roughly 1000 image prompts.
One important difference besides the price: the local version runs without restrictions, while the web version blocks prompts that might generate sexual or violent images, for example. DALL-E 2 and Midjourney also have this restriction.
Stable Diffusion: AI training for relatively little money
Training the image AI was relatively inexpensive, Emad Mostaque reveals on Twitter. The mathematician and computer scientist founded Stability AI, the startup that is the driving force behind Stable Diffusion.
According to Mostaque, the Stable Diffusion team used a cloud cluster with 256 Nvidia A100 GPUs for training. This required about 150,000 hours, which Mostaque says equates to a market price of about $600,000.
For DALL-E 2, Mostaque assumes a computational cost of approximately one million A100 hours. The training costs of OpenAI's image AI are thus likely to be far higher than those of Stable Diffusion. However, DALL-E 2 currently offers better performance, in part due to its architecture, which does require more training data.
Dalle2 end to end (including vit-h and upscalers) is probably 1m A100 hours. On the nvidia eos supercluster (5760 h100s) assuming 3x uplift could train in like 60 hours.
— Emad (@EMostaque) July 13, 2022
AI training for large models is affordable
Mostaque's explanation is interesting in two respects: First, he mentions concrete training costs. For other large AI models like DALL-E 2 or GPT-3, there is only speculation, sometimes in the millions, but no facts. Thanks to Stable Diffusion, there is now a concrete reference point.
On the other hand, the training costs of $600,000 are within a financial range that many companies can afford.
This, in turn, is an indication that a dominant position of companies such as OpenAI in the field of large AI models need not arise, at least, because of training costs that are unaffordable for others. The real cost drivers are likely to be research and development personnel and data collection and maintenance.