Extreme Finishing Time

[Rachelnowhere]

When using Automatic1111 with SDXL all images have an incredibly long finish time, so do other Checkpoints (like 1.5).

As seen above, the progress of the image appears to be finished in 0.32seconds (reflected the same in the UI progress slider which reads 97%) but the image won't appear in the output folder until 1minute and 06seconds. The progress slider is stalled at the very end of the image's generation.

Does anyone know why?

Additionally, the IMG2IMG function doesn't work until the TXT2IMG function has been used first. If I load A1111 and want to use IMG2IMG I need to generate a random IMG2IMG image. This isn't really an issue, but thought I'd mention it.

[Dagdgg]

No clue why but it's a problem with A111 for sure, it does not happen with ComfyUI nor Forge so there is something screwing with any of the functions related to saving the rendered image.

Having tried with or without a custom temporary folder I can clarify that the wait is there regardless.

I always get at least 2-3 seconds of wait, normally about 10sec and I have had instances like yours where it's taken up to... 2min 30sec or something near there.

Edit: Oh yes, and minimizing windows tends to help. Minimizing any open windows, particularily the webui browser window, or if you happen to have a youtube video running while generating can help it. It seems to be set on hold due to some external processes interfering which can come simply from the browser running the webui

[w-e-w]

sorry I made a mistake earlier
can you perform the profiling test again
but this time before launching webui
add set CUDA_LAUNCH_BLOCKING=1 after COMMANDLINE_ARGS in webui-user.bat

set COMMANDLINE_ARGS= .......
set CUDA_LAUNCH_BLOCKING=1

then launch webui and perform the test

remeber to CUDA_LAUNCH_BLOCKING afterwards
CUDA_LAUNCH_BLOCKING make webui performed slower then normal, so normally you wouldn't want to use it unless when doing profiling

[Dagdgg]

Sure sure, using the normal VAE and no-half-vae right?

https://www.dropbox.com/scl/fi/z9kmoma4tp78hz0ak7zh7/trace_NoHalf.json?rlkey=0lbq8c7154v0p7ilsam5m03h0&dl=0

Or with the nan-check?

https://www.dropbox.com/scl/fi/14o39thskb0vq9gtabs7w/trace_NoHalf_NoNaN.json?rlkey=9fh3pyy41jp5dqt1tqbro7zzu&dl=0

[w-e-w]

so after I sent the previous message about CUDA_LAUNCH_BLOCKING
I realized you sent a second message before I replied

and so I read about "... removing no-half-vae fixes the problem ..."
to me that's basically indicates that it's mostly a VRAM issue
basically what I said about above

also considering that you're you on the 12 GB card ...
...

• [Resolved] NVIDIA driver performance issues #11063

and it has nothing to do with nan check
--disable-nan-check and from the correct profile results we can see actually only takes and negligible ~200 µs
so there is no reason not to use allow nan check

I composed my reply but I forgot to press Reply and close the web page .....

---

whatever configuration you're using now if it's not slowing down for you it's good

remember about I suggest you use --medvram-sdxl because of vram
well ... I was just testing stuff and I might have stumble across a bug I'm not too sure with --medvram-sdxl
so I'm not sure if I should recommend you using --medvram-sdxl at this point

[Dagdgg]

Cheers, I have used --medvram plain and --medvram-sdxl and had the same problem earlier.

[w-e-w]

if it turns out cause of the issue is what I'm described then this extension might help
Tiled VAE from https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111
https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111/wiki/Tiled-VAE

It saves your VRAM at nearly no cost.
You may not need --lowvram or --medvram anymore.

---

also discard what I mention about the medvram bug
it turns out it was caused by a modification I made but I forgot revert to a code when I was testing something else
and it just so happens that thinkg doesn't play well with medvram

[thundercat71]

Do you have something like Restore Faces etc enabled? These will slow down the end generation as they kick in at the end.