Two months ago we moved our image generation architecture to leverage on serverless architecture where it allows us to remove the need for a queuing system when you send an image generation request.
Recently, we made some improvements and managed to shave around ~500ms on each image generation. It may not sound a lot but it is definitely a noticeable improvement if you’re generating thousands of images. On average you would save 8.333 minutes per 1000 images.
We are always on the look to improve this further in the near future.