lllyasviel
The system undergoes a downgrade after roughly every 100 generations.
I have noticed a significant downgrade in generations after running about a hundred generations. Currently, I have to uninstall approximately every 100 generations. The hands start to develop six fingers, malformed body anatomy, and strange faces. The problems seem to recur after some generations when someone appears nude. And it is not a single isolated episode. I have experienced this more than 10 times already. Does anyone have any quick solutions besides reinstalling?
I haven't been experiencing the same kind of issue with Fooocus. Can you confirm that generated images change after 100x using the same settings with the same seed? Can you please add step by step instructions to reproduce this? Thanks.
I haven't been experiencing the same kind of issue with Fooocus. Can you confirm that generated images change after 100x using the same settings with the same seed? Can you please add step by step instructions to reproduce this? Thanks.
I have noticed that the prompts start to be less responsive to some degree, and the images generated using the same prompt in an older installation appear more deformed and contain anatomy errors, while in the new installations, prompts seem to be better understood and yield results with proper anatomy and structure. Is there any type of user cache or tool in focus on learning user preferences that could be causing this after extended usage?
If there is, do we have a way to clear that easily without needing to reinstall Fooocus and all its components?
I haven't been experiencing the same kind of issue with Fooocus. Can you confirm that generated images change after 100x using the same settings with the same seed? Can you please add step by step instructions to reproduce this? Thanks.
I have noticed that the prompts start to be less responsive to some degree, and the images generated using the same prompt in an older installation appear more deformed and contain anatomy errors, while in the new installations, prompts seem to be better understood and yield results with proper anatomy and structure. Is there any type of user cache or tool in focus on learning user preferences that could be causing this after extended usage?
If there is, do we have a way to clear that easily without needing to reinstall Fooocus and all its components?
I tried using same prompts and negative prompts too.
@MoacirEduardo i can't reproduce this but maybe somebody else can confirm.
I have generated a ton more iterations, and I've noticed that when making a lot of iterations in a single session, memory usage starts to increase, and average precision begins to decline. That's why I asked if there is any type of cache or user preference data stored in the Fooocus files after prolonged use. If so, I'd like to know if there's a simple way to clear it for each new extended usage session.
I am not aware of any cache or any way to clear it, but maybe there is a memory leak. Related to https://github.com/lllyasviel/Fooocus/issues/1646
I have noticed on my computer very similar phenomena. In my case, instead of a deterioration in quality, problems begin to appear with the generation of human body anatomy and other objects despite using the same parameters all the time. Some objects in generated images or inpaint are definitely degenerated no matter what checkpoint I use. This will be the 3rd reinstallation since starting to buy a new computer. Updating graphics drivers just for your information does not help.
I have noticed on my computer very similar phenomena. In my case, instead of a deterioration in quality, problems begin to appear with the generation of human body anatomy and other objects despite using the same parameters all the time. Some objects in generated images or inpaint are definitely degenerated no matter what checkpoint I use. This will be the 3rd reinstallation since starting to buy a new computer. Updating graphics drivers just for your information does not help.
That's exactly it. You explained this very same problem I had in a concise manner.
So it is indeed reproducible. The question is how many generations exactly do we have to make to achieve this state in Fooocus. I also tried yesterday to repeat it using webui and didn't manage to get anything like it (I generated 200 images on a clear installation, and at the moment everything looks good so far).
I think this is possibly what I am experiencing. I have probably over 100s of images generated at this point, but I notice the prompt is almost ignored. I have tried weights (come from ComfyUI background) and it is extremely unwieldy, rigid. The output is almost equivalent to stock images, you can't vary anything and using a one word prompt seems to result in the best images (but nonetheless not desired!)
