Older versions could run forever without any problems, and had a hi-res fix that didn't spit out 5 copies of the image. I could cue up 200+ images in one go. Now I can't even que up 3 without getting a CUDA out of memory error. Meanwhile Auto1111 can do continuous generation on the same machine with no problems. Older versions of invoke also didn't bloat the .cache folder with piles of useless copies of the models every time a model is switched. The fact that I have to go in and manually delete that junk data is absurd. Invoke has become continuously less convenient to use, and can't even automatically load models currently. What is going on?
#It really feels like Invoke has backtracked significantly in the last year.
bleak fiber
thick arrow
Does number of images in queue actually impact anything performance-wise? I would've thought only image size and stepcount would matter?
bleak fiber
If I ran one at a time, I could keep generating indefinitely, if I ran 3 or more, CUDA out of memory would happen within a few minutes consistently, requiring closing invoke, emptying cuda cache, and restarting
On a fresh install of 3.6, default settings
orchid gust
Would like to see some stats and configs for your system - this is atypical
If you could share your config file, and then the output after the generations running into the OOM error, that'll help us help you.
bleak fiber
It's a 4gb vram laptop that experiences this issue. It ran older versions of Invoke (2 for example) without encountering Out of Mana errors
I'm aware 4gb vram is low for image generation
orchid gust
This may just be you needing to configure invoke as a result of new options we’ve added
Setting your model caches to 0, enabling the performance optimizations, etc
bleak fiber
Setting your model caches to 0, enabling the performance optimizations, etc
I have done those things, I learned in version 2 that having the caches above 0 wasted resources and slowed down the process. Regardless, this has nothing to do with the particular settings I happen to have, but how the program functions in particular states. How the CUDA Out of Memory error is arrived at is irrelevant. The point is, when it happens, emptying the CUDA cache fixes it.
orchid gust
You might need to share more details - might be an edge case since you’re on such a tight card