Sampling scheduler
After discussing the generation seed, let’s now delve into another essential aspect of Stable Diffusion image generation: the sampling scheduler.
The original Diffusion models have demonstrated impressive results in generating images. However, one drawback is the slow reverse-denoising process, which typically requires 1,000 steps to transform a random noise data space into a coherent image (specifically, latent data space, a concept we will explore further in Chapter 4). This lengthy process can be burdensome.
To shorten the image generation process, several solutions have been brought out by researchers. The idea is simple: instead of denoising 1,000 steps, what if we could take a sample and only perform the key steps on that sample? And this idea works. Samplers or schedulers enable the Diffusion model to generate an image in a mere 20 steps!
In the Hugging Face Diffusers package, these helpful components are referred to as schedulers. However...
