Hey folks, just getting started with Stable Diffusion and trying things out. I installed the AUTOMATIC1111 web UI locally on my machine, which has a decent GPU.
But wow, generating a single 512x512 image feels like it takes forever! I see people online talking about batches and speed, but for me it's like 30 seconds per image. Is that normal?
Am I missing a setting, or is the web UI just kind of slow by default? My GPU usage doesn't even seem to max out. Feels like there's a bottleneck somewhere. 😅
Maybe my expectations are just too high from seeing all the fast demos?
Thirty seconds for a 512x512 image is definitely on the slower side for a "decent GPU." The web UI itself isn't inherently that slow; it's more likely a configuration or bottleneck issue.
Your observation about GPU usage not maxing out is the key. The most common culprit is running in --precision full, or fp32 mode, instead of --no-half or similar flags that force full precision. This dramatically slows things down without fully utilizing the GPU's tensor cores. Also, are you using an Nvidia card? If you're on a laptop or have an AMD GPU, there are different optimization paths.
Could you share your command line arguments and GPU model? That would help pinpoint if it's a precision issue, a VRAM limitation causing constant swapping, or something else like a slow model loading from disk.
You're not wrong. The web UI's default setup is inefficient.
> My GPU usage doesn't even seem to max out
That's the real problem. They don't ship it optimized. You have to manually set the right flags for your hardware, otherwise it's running in a slower mode that doesn't fully use the GPU. It's a common complaint buried in the issues on GitHub.
People showing off fast demos are using those optimized launch arguments. The default install gives you the slowest, most compatible settings.
Your point about GPU usage is correct, but "configuration or bottleneck" undersells the core problem. The defaults are a compatibility sandpit that guarantees a bad experience for performance. It's not just about the user finding the right flags, it's about the tool shipping with deliberately poor defaults.
Forcing full precision is the main culprit, as you said, but the other big one is if the user is on a CPU with older instruction sets and doesn't have the right arguments for that. The load times and sampling can crawl. It's a data quality issue for the setup process, basically failing to profile the user's system and apply sensible presets.
People shouldn't need to trawl GitHub issues to get baseline performance. The demos set an expectation the default install doesn't meet.