-
Notifications
You must be signed in to change notification settings - Fork 3.9k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
After the last update black images after few steps on Mac mini M2 pro with Sonoma 14.5 #3521
Comments
I have the exact same issue here. |
Can you tell me which checkpoint and workflow you are using? |
Happen with all checkpoint |
If you run it with |
Here in Italy is 01:00 night, so I've made only one test and it works with that fix, but Comfyui was very very slow to start, but it is a begin to discover a better solution. |
I have same issue on mac m3 pro sonoma. but use --force-upcast-attention can fix the issue |
Hi everyone, |
If you have the issue can you post your pytorch version here, you can print it with: |
2.4.0.dev20240518 |
If you install stable 2.3 is the issue present? |
I am not sure it's appropriate to do that, as Apple recommends the nightly version. I will probably try but so far, using --force-upcast-attention has resolved the issue. |
Mine is :2.4.0.dev20240413 |
Noob here. Where do I enter --force-upcast-attention ? I'm on a Macbook. Do I enter it in terminal by itself or in addition to python3 main.py ? |
Possibly related: #3523 Will try the suggested fix tomorrow |
you are adding --force-upcast-attention after python3 main.py |
为什么我的是:main.py: error: unrecognized arguments: --force-upcast-attention |
2.4.0.dev20240521 on M3 dpmpp samplers (with and without karras) caused a black image. The |
update comfyui? |
没有update comfyui,因为更新到最新版后采样器几乎全部失效,而且用ComfyUI Manager 这个插件,只要选择将Preview method:更改为Auto,结果运行生图就会出现以下错误,python意外退出。Python(5424,0x16b8ef000) malloc: Incorrect checksum for freed object 0x33c1a4c00: probably modified after being freed. |
Same issue, almost all workflows produce black frames. (OS 14.5 on M3 Max) Starting comfyUI with python3 main.py --cpu seems to solve it, but it's very very slow |
The problem is not only on Sonoma, it is also on Ventura. It starts in May with some ksampler before being extended to all.
Exactly Karras now seems to be very slow now
Read what I wrote in the issue #3567 |
If you update ComfyUI the issue should be solved. |
Black Image again. --force-upcast-attention solve the problem yesterday but now the problem is here again after the update |
The problem arrive if the batch size is > 1 with --force-upcast-attention everything is ok |
I am on a Mac M2 Max with 64 gig ram. Same issue. --force-upcast-attention doesn't do anything to help. Comfy has always been very reliable and stable on my machine, now I can't even generate using SD1.5 models. It's either black or gobbledegook depending on the sampler I choose. |
Thank you [comfyanonymous], I'm on MacOS 14.5 with M3Max. --cpu did not generate black frames but was painfully slow. Non workable. I tried to load an archived folder of ComfyUI, before my IPAdapter update, but it didn't work. I tried removing the entire ComfyUI directory and reinstalling fresh ComfyUI, also didn't work. With the help of ChatGPT4o, I uninstalled home-brew and removed all versions of python I could find. Re-installed Home-brew, Python 3.11.9 (also verifying that this is what's active and it's not calling an earlier version when I start Comfy) and latest Pytorch Nighly. Took me 3 days to troubleshoot, but now everything is working as it did before with my full ComfyUI directory. |
Update- I had luck by updating torch to 2.3, now everything seems to be working again? |
Everything seems to work smoothly again. A bit slower though... |
It is repair thanks. No need --force-upcast-attention |
The problem starts again. I don't explain why. |
After update, any execution took almost x2 time than before. When steps >= 8, black image appear again. Torch ver: 2.4.0.dev20240526 |
FWIW, same issues for me.
Model: Macbook Pro, 14-inch, 2023 $ python -c 'import torch;print(torch.version.__version__)'
2.3.0.dev20231215 |
Unexpected value in cast, It seems to be linked to denoise and seed number. Persisted after update but deleting samplers from earlier workflows and replacing them seemed to fix it for now. No SD1.5 models have been working and very few SDXL ones. I updated everything today PyTorch etc. |
Further to my last comment for me it only happens on smallish (768 or 1024) square images, Ultimate upscale fails every time with black tiles everywhere. |
The first problem you mention seems to occur before the black screen for me too. It looks like Ksampler is sometimes corrupted we have to replace it. |
It happens partway through the sampling, at about step 6 or whatever it goes black. It doesn't always report the cast error which is very strange. By changing denoise and steps I can get it to work again and once going it won't happen again on that seed. |
my error is ComfyUI/custom_nodes/ComfyUI-VideoHelperSuite/videohelpersuite/nodes.py:95: RuntimeWarning: invalid value encountered in cast |
How wrote in the subject, almost every workflows not working, I see the preview, but after few steps the image goes black.
After many test I've discovered that "Karras" scheduler was causing this problem.
The text was updated successfully, but these errors were encountered: