Closed Gouden89 closed 1 year ago
Thank you for the report. As far as I can tell, the NansException is related to the model, lora, vae, embedding, and has nothing to do with the extensions. I've seen some of them throw a NansException, just on inpainting.
I can see in your prompt that you have a lot of highly weighted things like <lora:add_detail:1>, <lyco:GoodHands-beta2:1.0>
. One of the tips from the Korean community is to keep the sum of the weights of all your LORAs within 1.5. This will help to prevent NansException.
Or, you can put prompts like "best quality, masterpiece" in ad_prompt to prevent inpaint from using loras.
tried with a different model and a simple prompt "a girl sitting in a chair", no loras or anything and get the same error. Works fine without adetailer enabled and with it bugs out after the first generation
Check out this: https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/6923
There are so many things involved in this issue that it's actually hard for us. sorry.
I created a PR for A1111 that fixes a similar problem with the Controlnet extension. https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/11569
You can give it a try and see if it solves the issue.
Also, make sure that persistent cond cache
is unchecked under Settings -> Optimizations
@ramyma Thank you. I think it could help me troubleshoot.
I'm also having this same issue. My only LoRa is at a weight of 0.5, but generation fails whenever I go for batch counts more than 1.
I can also confirm that everything works fine so long as ControlNet isn't used with ADetailer. Using tiled or inpaint causes the error with multiple batches - though I haven't tested with other ControlNet models.
@Neverdusk Thanks for the report, but can you post the terminal logs as well?
@Neverdusk Thanks for the report, but can you post the terminal logs as well?
Sure - should I post them straight into here or is there a better way to hand them over? In case I can avoid bloating the thread with a long log.
As one more point (I'm still testing different idea) - I just figured out that the error CAN be avoided, even if ControlNet is enabled - if the FIRST ADetailer does not use it. I managed to run generation with a batch size of 2, with the first ADetailer (person8n-seg) not having any ControlNet but the second ADetailer (face8n) using Inpaint Global Harmonious, and it worked perfectly.
However, in a previous test, adding ControlNet to the first ADetailer but not the second still caused the issue. So it's fine if ControlNet is used at the end of a batch / between batches, so long as it isn't used mid-batch.
I can edit your comment, so just post it here. Except if you're viewing on a mobile device, it's fine.
Here's a fresh log. I created the same two images (batch size 2) repeatedly with ADetailer enabled. On the first run, neither used ControlNet. On the second run, the second ADetailer used ControlNet. On the third run, both used ControlNet, and generation failed:
The latest version of WEBUI, there will be this problem when using controlnet in I2I, please help to solve it, I can only use Face Editor to replace this plug-in
I've noticed this same problem with batch sizes greater than 1. When I run the exact same batches with adetailer turned off, they come out ok.
Describe the bug
So before I could do higher batchcounts but since a couple days I get this error. Without adetailer enabled it does work, but with it generates 1 picture and then it bugs out.
The entire console log, including python version information, webui version, commit hash, commandline arguments and error traces.
List of installed extensions
No response