How does the Stable Diffusion algorithm work?

06 Tem 10:30

Last month I got stuck on a beast of a project, honestly just using Stable Diffusion without really knowing what was going on under the hood. Then it finally clicked for me—like when you see how the noise reduction part works so seamlessly that when your images render weirdly, you realize the latent space and the diffusion space are constantly feeding off each other. That "oh my god, I finally get it" moment was when I realized that while generating an image, I’m starting from this tiny latent vector and gradually, step by step, like exorcising a demon, stripping away the noise—kind of like using Photoshop’s tools to tweak details layer by layer. To put it even clearer: the diffusion process starts with your image looking like a fragmented mess, like the *Communist Manifesto* of noise. Then the U-Net model slowly cleans it up with each step, pulling it toward the image you’ve stored in the latent space. So the latent space is basically like a "USB drive holding the summary of your image"—each noise reduction step makes that summary sharper. When I saw the code line `noise_pred = unet(latent_model_input, timestep)`, it hit me: the latent space and the diffusion process are dancing together like identical twins.