There's currently an issue with GTX 1660 series cards that SD will generate black images unless it's set to full precision. There doesn't seem to be a way with this tool to run SD with the '--precision full' option.
There been a few reports like this. It seen that 1660 need to run on full precision, but not sure if 6vram will able to run a good resolution at full precision.
It seems no matter what I do, all of my results are just black images. I've tried different steps from 50-100. Tried 64, 256, and 512 resolution. But regardless. Nothing works.
Hi there. It use 1.4 I'm not sure about this txt2img, I did my own scripts on top of the huggingface code and my own optimizations, so I don't know how it compare to the txt2img
Thanks ! i was able to fix it by just downloading the Rar again and Setting the resolution to 512x512, i have a RTX 2070, maybe some of these infos will help people who have similar issues :)
torchvision\io\image.py:13: UserWarning: Failed to load image Python extension:
torch\_jit_internal.py:751: UserWarning: Unable to retrieve source for @torch.jit._overload function: <function _DenseLayer.forward at 0x000001C305192700>.
warnings.warn(f"Unable to retrieve source for @torch.jit._overload function: {func}.")
torch\_jit_internal.py:751: UserWarning: Unable to retrieve source for @torch.jit._overload function: <function _DenseLayer.forward at 0x000001C3051A7A60>.
warnings.warn(f"Unable to retrieve source for @torch.jit._overload function: {func}.")
"
It then loads fine, runs, but produces black images as output. Any ideas?
This errors at the start are normal for this first version (0.1) But it should not be producing black images, you are the second person with this problem. Can you check if you have all the read/write permissions on the folder?
I was originally getting permission errors when trying to save the image. I resolved that by adding exceptions (which worked), I even turned of my antivirus completely. The images get saved fine, it's just that they are uniform black. If anyone figures it out, please let us know.
Nope, no overclocking done. I tried some different settings of my video card to no avail. I also don't get any errors in the CMD window except for the normal ones at startup described by GRisk.
← Return to tool
Comments
Log in with itch.io to leave a comment.
There's currently an issue with GTX 1660 series cards that SD will generate black images unless it's set to full precision. There doesn't seem to be a way with this tool to run SD with the '--precision full' option.
There been a few reports like this. It seen that 1660 need to run on full precision, but not sure if 6vram will able to run a good resolution at full precision.
The 1650 Super does too, just a heads up
Do you have a solution for enable a multi GPU VRAM render, I have 3x3090 who want to burn some pict...
In the future, only to speed up samples >= 3, it would be very hard to use 3 cards to generate a high resolution image.
Ok, at this time with a 3090 I can generate a 768x768 pixel picture,
Thanks for your work.
Nothing happens when I open Stable Diffusion GRisk GUI.exe, nothing opens. What should I do differently?
Can you open CMD and then open the .exe with it? This way the windows will not close and you can send me the error
do you plan on implementing image-to-image generation?
Yep
Can't wait, I run it in google colab and it's annoying to enter everytime and install stuff
thats good to hear, cant wait! img2img is something ive always wanted to try with stablediffusion.
Something that's just click-and-go. That was quick, thanks.
It seems no matter what I do, all of my results are just black images. I've tried different steps from 50-100. Tried 64, 256, and 512 resolution. But regardless. Nothing works.
Are you using a 1660 card? It seen to be a bug in it with half precision models.
Aw yeah I am. Sadly when clicking the Half-Precision mark, nothing appears to work...
Tried 150it and 256x256 with one prompt. Keeps outputing black screens. What could I be doing wrong?
512x512 is what SD likes most, anything else and it doesn't have a good chance of working
Yeah, I tried that and get a CUDA out of memory error. Is there a work around? I have more than enough shared memory.
GRisk you are a god
why work only Nvidia? AMD?
pytorch is made mostly for CUDA(Nvidia)
AMD can work in the future, but it's some extra work
nice, thanks for reply, regards
Hi, firstable thx for this release.
Which model do you use ? 1.3 or 1.4 ?
Is it the optimized txt2img.py or the official one ?
Keep on the good work !
Hi there. It use 1.4
I'm not sure about this txt2img, I did my own scripts on top of the huggingface code and my own optimizations, so I don't know how it compare to the txt2img
Thanks for the answer and this 0.1 version.
index 1000 is out of bounds for dimension 0 with size 1000
any idea what makes this happen? i ran it like 50 times before and then suddenly it stopped working :(
It need to be multiple of 64. So try rendering 1024X1024
Your card also need to be able to render this much res. 10vram maybe can do it, but not sure
My RTX 3080 FE 10gb VRAM card was unable to do 1024X1024
It may work with 960X960 but you will need to keep testing until it work
Thanks ! i was able to fix it by just downloading the Rar again and Setting the resolution to 512x512, i have a RTX 2070, maybe some of these infos will help people who have similar issues :)
for me it happened because of too high sample count (works below ~250)
You should include the version number in the .rar name. Thanks for making this!
Good idea. Would need to re-upload for that tho, so gonna do it in the next update.
Works very well! Thank you very much. I wonder if there are any plans to add image to image as well.
Yeah, img2img will appear in the future. Still a lot of work to be done.
I keep getting the below error when loading:
"
torchvision\io\image.py:13: UserWarning: Failed to load image Python extension:
torch\_jit_internal.py:751: UserWarning: Unable to retrieve source for @torch.jit._overload function: <function _DenseLayer.forward at 0x000001C305192700>.
warnings.warn(f"Unable to retrieve source for @torch.jit._overload function: {func}.")
torch\_jit_internal.py:751: UserWarning: Unable to retrieve source for @torch.jit._overload function: <function _DenseLayer.forward at 0x000001C3051A7A60>.
warnings.warn(f"Unable to retrieve source for @torch.jit._overload function: {func}.")
"
It then loads fine, runs, but produces black images as output. Any ideas?
This errors at the start are normal for this first version (0.1)
But it should not be producing black images, you are the second person with this problem. Can you check if you have all the read/write permissions on the folder?
I'm having the same issue. Has any solution been figured out yet?
Are you by any chance using a notebook or have two cards in the computer? It may be selecting the wrong card.
How would I go about verifying that the right card is being selected? I've never used Stable Diffusion prior to this, so any advice would be helpful.
I was originally getting permission errors when trying to save the image. I resolved that by adding exceptions (which worked), I even turned of my antivirus completely. The images get saved fine, it's just that they are uniform black. If anyone figures it out, please let us know.
Same here, WIN 10 - AMD Ryzen 7 2700 - 32GB - NVIDIA GeForce 1660 - 6GB - Driver 516.94
Result is solid black png 512x512.
If you have Overclocked your GPU that could be the problem, try putting it back to its defaults.
Nope, no overclocking done. I tried some different settings of my video card to no avail. I also don't get any errors in the CMD window except for the normal ones at startup described by GRisk.