Millist tüüpi U-Net kasutatakse Stable Diffusion'is? Millised on selle omadused võrreldes tavalise U-Net-iga?

In Stable Diffusion, a modified version of the U-Net architecture is used, adapted for image diffusion modeling tasks.

Features of this U-Net compared to the classic:

• Input data: Instead of regular images, a noisy image is provided along with a diffusion time step (noise level), which is encoded and fed into the network.
• Time injection: Mechanisms are added to the architecture to account for the time step (time embedding), allowing the model to consider the degree of noise.
• Use of attention: Some layers of U-Net incorporate attention mechanisms (self-attention), improving the model's ability to capture global dependencies in the image.
• Modified blocks: Instead of standard convolutional blocks, ResNet blocks with normalization and activation are used, optimized for stable training.

Thus, U-Net in Stable Diffusion is a specialized architecture that combines the classic structure with additional mechanisms to work with noise and temporal steps, which is critical for diffusion models of image generation.