GFPGAN causes VRAM leak with batches about stable-diffusion-webui

[AUTOMATIC1111]

That's strange. I'm not getting any leaks. Task Manager's GPU memory usage says pretty much the same. I generated 10 batches with 4 images each, vram usage goes to 22.7 when images are generated then to about 11 i guess that's when GFPGAN is working, and then 10.6.

[hentailord85ez]

I will add that I am on Linux (Nvidia official drivers), however I don't assume that to be the problem as Python & co. are platform agnostic.

Although official Linux instructions are not in the README, just following the python-specific steps should give the same outcome. The only difference I would imagine is not being able to use the .cmd .

The last message in issue #9 (after it was closed) also mentioned of the bug happening after the fix PR. I assume that they are not using Linux too, although I might be wrong and it is just a Linux specific issue.

My settings are vanilla, I've only changed the sliders provided in the text2img & img2txt panes.

I've attached a screenshot of the output of nvtop while generating images.

Green line is compute utilization %, orange is VRAM%.

Edit: I have tested GFPGAN functionality on that certain other UI fork, and it does not leak VRAM, which eases the fear that it's my shoddy installation of Linux.

[orionaskatu]

I'm also on Linux (RTX 2070 8GB) and I have a similar issue, tested on a batch count of 9.
Without GFPGAN VRAM usage is stable, but with GFPGAN it keeps growing.
I don't get an OOM but the generation is stuck after the 6th image and VRAM is not cleared.
I'll try to do more tests to figure out what's going on.

For the record I launch with theses params to prevent OOMs --lowvram --always-batch-cond-uncond and I use my own fork (very slightly modified, I'll try with the original webui).

Edit: With the unmodified webui.py I get an OOM after the 6th image with GFPGAN.

Edit 2: Without --always-batch-cond-uncond it uses a lot less VRAM but I still get an OOM after few images with increased resolution. I'll try --medvram and without options.

Edit 3: Same issue with --medvram and no options, I'll try on Windows tomorrow if I can.

[AUTOMATIC1111]

Let's see if this changes anything.

[orionaskatu]

You nailed it!
The VRAM usage is now stable and no more OOMs, thanks !

[AUTOMATIC1111]

This is upsetting. This means GFPGAN leaks memory if you create one, dump it, and then create another. I was hoping to not keep it in memory for low VRAM users.

[hentailord85ez]

Just tested it; it's stable for me too. Thanks for the the fix :)