The Hidden Variables: Models, Memory, and Hardware
AUTOMATIC1111/stable-diffusion-webui does not run in a hardware or model vacuum. The checkpoint, VAE, device backend, precision path, and memory strategy all condition the pipeline that turns a prompt into an image. The practical focus here is how to identify those variables and choose among resource strategies without reducing compatibility to a simple question of whether a machine “supports” Stable Diffusion.
A prompt and seed are meaningful only inside that surrounding state. The same recorded request can behave differently when it is sent to another checkpoint, paired with another VAE, executed through another precision path, or moved from an accelerator to a CPU-oriented path. AUTOMATIC1111 makes many of these variables selectable and records several of them, which is useful for comparison—but it also means that a working setup is a composition that must be understood and maintained.
A checkpoint is an identity, not just a file
A checkpoint contains learned weights that give the model its behavior. AUTOMATIC1111 scans for checkpoint files, reads metadata from supported formats, calculates identities, and detects structural families such as SD1, SD2, SDXL, SSD, and SD3. That detection is not a cosmetic label. The model family determines which internal components are present and which processing paths can reasonably be applied. A feature exposed by the interface may still be unavailable or unsuitable for a particular model, and the presence of a family label does not establish that every external checkpoint is compatible with every sampler, extension, or auxiliary model.
The VAE belongs in the same discussion. It participates in the conversion between a model’s latent representation and a visible image, and the runtime can select a VAE separately or resolve one associated with a checkpoint. Two runs that appear to use the same prompt and sampler can therefore differ if their VAE state differs. The model choice and the VAE choice are not incidental loading details; they are part of what the processing pipeline means by “decode this result.”
Hashes make the identity problem more concrete. The runtime can attach short and SHA-256-derived identities to checkpoints, read metadata from safetensors files, and place model or VAE names and hashes in generation information. A filename such as portrait-model is a human label that can be changed or duplicated. A hash gives a stronger handle for asking whether two runs used the same artifact. It does not certify the artifact’s quality or origin, but it helps prevent an operator from comparing two files as if they were one.
This matters most when a result changes unexpectedly. Suppose a prompt, seed, sampler, and image size are copied from an earlier result, but the new output has a different composition. Before changing the prompt, check whether the checkpoint identity and VAE match. If they do, the next questions concern the processing settings, extension state, precision path, and device. If they do not, the model boundary is already a sufficient explanation for why the same recipe no longer means the same thing.