Stable Diffusion

The original Stable Diffusion repository describes a particular v1 architecture built on latent diffusion research. An autoencoder maps image pixels into a spatial latent representation, a U-Net carries out the denoising predictions, and an associated decoder maps a sampled latent back into pixels. The repository specifies a downsampling factor of eight. For a 512-by-512 image, dividing each spatial axis by eight gives a 64-by-64 latent grid; this is a spatial count rather than a statement that every latent has only one channel. Text conditioning in v1 uses a frozen CLIP text encoder and non-pooled text embeddings connected through cross-attention. The prompt therefore affects the learned denoising process rather than simply setting a filename or being printed literally onto the image. Classifier-free guidance combines conditional and unconditional predictions using a chosen scale. Its effect should be evaluated rather than described as a guarantee that larger scales always improve fidelity or prompt adherence. Sampling algorithms and step counts change the computation and output, with quality depending on the checkpoint, solver, and settings. More steps are not universally better, particularly when using a model intended for a different sampling regime. Published weights and code permit technical inspection and supported extension, subject to their stated license and model constraints; open weights do not mean unrestricted rights or compatibility with every device. ControlNet studies adding spatial conditions, including edge, pose, and depth information, to pretrained diffusion models. Such an extension is another component and does not belong automatically to every original v1 checkpoint.