I have a 308010Gb but I can't use this software, at start, the windows command line say some errors:
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 0x0000017B06DDD550>.
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 0x0000017B06DF38B0>.
warnings.warn(f"Unable to retrieve source for @torch.jit._overload function: {func}.")
Thanks, I search by the word "torch" in this webpage and have 0 results when I posted, so i suspect they added the information after I post this question.
I've created 704x704 images, but it is quite a bit slower. (1.20s/it vs 3.5it/s on 512) Do you really see a difference in coherency? maybe I need to try a little bit more.
Seems that you just keep going and get enough single heads. My prompts are "sticky": if I change a prompt, it seems to keep the previous prompt for several photos, and then moves in the direction the prompt I asked it ... !
what I find best is to set steps to a low number like 20 or 30 then generate like 100 images with the same prompt but random seed. Delete the bad ones and keep the ones with the right general composition then use the seed from it and iterate on the prompt using a higher number of steps.
what has worked for me is when I get a render I like, I pull the seed out of the _info, and set it, then I'll up the iterations to 150. then i'll slowly start increasing the v scale 1 at a time (somewhere around 13-14 comes out best). Thankfully I have a 3090, so each 512x1024 only takes 90 seconds.
Thanks for creating this. I'm enjoying creating privately vs. Midjourney's public forum. Is there a way to point the program to a GPU1. I'd like to see how it performs on the other GPU in my system. Thanks!
So Stable Diffusion is what Dream Studio uses, Stability.ai developed both the model and the service.
Midjourney uses something else, but did recently run a short test integrating Stable Diffusion into their service, whatever their implementation of it was, it worked very well, but it's offline for tweaks right now.
Working perfectly on my RTX 3060(12GB) at about 3.47it/s, but with a hard limit at 512 of resolution and 250 steps. It's awesome nonetheless. Thank you for your quick and awesome work.
Edit: I've been able to get 704x704 images but they go at 1.15s/it. It's a lot slower, not worth it for me.
Did you try turning on the "Use Half-Precision" ? This makes it possible to be used in hardware that is less than 10GB of VRAM, maybe with this enabled you can achieve grater resolutions and/or speeds because it will have less precision, so less vram is needed. (enabling this will do float16 instead of float32).
If you want more speed, I guess, use the same resolutions with half-precision enabled... I cant test this myself, because I only have 8GB VRAM, and GUI dont let me uncheck it
Does it work on a GTX 1650? Because it only lets my do 62 x 62 and under and even then it just makes a black screen. Does CUDA have to be installed to make this work because its giving me a CUDA memory error when I'm running it as 512 x 512, and I don't know why.
Is there any way to use this without needing an Nvidia driver? Every time I try to generate an image it tells me: RuntimeError: Found no NVIDIA driver on your system. Please check that you have an NVIDIA GPU and installed a driver. If not then that's fine.
this program only works with Nvidia graphics cards so that could be the problem, if you have an Nvidia gpu you can try to install the latest drivers with geforce experience and that might fix it.
When I generate an image at 150 steps with a resolution of 512X512 it only shows a black screen. I don't know if it's my GPU (RTX 2060) but it won't work.
Use google collab or try to install it without the GUI. This GUI is good but the default and only option of "half precision" makes it impossible to generate images in some nvidia graphic cards (GTX 1650, 1660, including SUPER) it happened to me, the solution was installing the code locally without any GUI, there are some websites that explain it:
Once you have it running, in the terminal use the command --precision full. That should solve the issue it did it for me! good luck out there and wait for the new version of this GUI this is just the 0.1v
I have a 1650 and my images were showing up green. Using the "--precision full" argument enabled the output, but I'm not sure how you could do that with this build
I’m not much for legal jargon. I’ve read the full license, when you state “You may re-distribute the weights and use the model commercially and/or as a service. If you do, please be aware you have to include the same use restrictions as the ones in the license and share a copy of the CreativeML OpenRAIL-M to all your users (please read the license entirely and carefully)”.
Does this pertain to the produced images being sold or the AI model itself being used by users of the GUI such as myself or others? Like do I need to put a link in my descriptions for craft supplies that this was used and put the copy of the license in there?
This license is a copy from the Stable Diffusion team, this line is just about the AI model. Images generated by the model are explained on the second line of the license. You can use the generated images as you like, no credit needed, but if you have more questions about the license, you need to talk with the Stable Diffusion team.
Thank you so much! I’m really excited for what this does for my use case. Looking forward to where it goes from here and I appreciate all of the work done to make a GUI and contained executable.
This is one of the coolest shit i seen, i been just playing around all night trying to get cool results, thanks a lot for soing this! i been trying to follow the guides from github but my non-programer brain has been strugling to get it to work, this works awesomelly!
Nice work, hopefuly future models / algorithm will allow for better output resolution thant 512x512 on a RTX 2070 Super. 8GB seems to be a little short for that ? (that's kinda crazy)
I have the same issue - happened after I chose to generate 4 ideas at a time. After that if I switch to just 1 it still tries to generate 4 ideas total where the first one is what you asked for and the rest is random stuff unrelated to the promp
I had the same issue, I realized I had an extra line without text so it would generate my initial prompt then move onto the second line, since theres no text it would just generate an incoherent image.
Ok thanks! Are you anticipating enabling full precision in a future update? If that's technically possible, I'd love to have the option to run Stable Diffusion through this program even if it takes a long time because of the amount of memory required.
I signed up here specifically to thank you! It works easy with my 3090, i love that the promt text stays after generating, very easy to work on the text !
Would be good if you could simply click on a few different resolutions, instead re-entering manually.
I expected you would be one of the first to make a GUI for this, good job! A great feature for a future update would be the ability to import an image and perform the AI transformations on that, some other GUIs have that feature so I imagine it is possible to implement
← Return to tool
Comments
Log in with itch.io to leave a comment.
I have a 308010Gb but I can't use this software, at start, the windows command line say some errors:
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 0x0000017B06DDD550>.
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 0x0000017B06DF38B0>.
warnings.warn(f"Unable to retrieve source for @torch.jit._overload function: {func}.")
Apparently if you read above, this always happens, and is of no consequence. So leave it and enjoy making pictures!
Thanks, I search by the word "torch" in this webpage and have 0 results when I posted, so i suspect they added the information after I post this question.
Anyway, it works fine, thanks!
this was last updated 3 days ago so it was there lol.
I managed to create 640 x 704 images with RTX 2080ti 11gb and coherency seems to improve from 512 x 512.
I've created 704x704 images, but it is quite a bit slower. (1.20s/it vs 3.5it/s on 512) Do you really see a difference in coherency? maybe I need to try a little bit more.
I have 4 gb vram but it says cuda out of memory and uses only 3.29 gb vram
With 4GB VRAM you'll need to run a lower image resolution like 384x384
512x384 will run with 4GB
Hi, I am getting double or triple faces in my pictures when I put "woman"/"girl" or similar words in my prompts to get portrait pictures.
How do I minimize getting these odd results?
Seems that you just keep going and get enough single heads. My prompts are "sticky": if I change a prompt, it seems to keep the previous prompt for several photos, and then moves in the direction the prompt I asked it ... !
Yeah, it feels like the more I generate the less I get double faces/persons coming up.
what I find best is to set steps to a low number like 20 or 30 then generate like 100 images with the same prompt but random seed. Delete the bad ones and keep the ones with the right general composition then use the seed from it and iterate on the prompt using a higher number of steps.
Awesome stuff, works out of the box with my GTX 1070 (8GB) at 1.3it/s. Wow!
1. What does V-scale do?
2. How do we make it use other models? (Can we do that yet?)
Higher and higher V Scale seems to give artefacts - I don't think its a fully-fledged option yet?
v scale is supposed to make the result more similar to the prompt but it just seems to curse the image from what I've tested. I like 5.5 or 6 though
what has worked for me is when I get a render I like, I pull the seed out of the _info, and set it, then I'll up the iterations to 150. then i'll slowly start increasing the v scale 1 at a time (somewhere around 13-14 comes out best). Thankfully I have a 3090, so each 512x1024 only takes 90 seconds.
Thanks for creating this. I'm enjoying creating privately vs. Midjourney's public forum. Is there a way to point the program to a GPU1. I'd like to see how it performs on the other GPU in my system. Thanks!
Is this a cousin of MidJourney or Dream Studio? I get confused when posting on facebook the "genealogy" of this software?!
So Stable Diffusion is what Dream Studio uses, Stability.ai developed both the model and the service.
Midjourney uses something else, but did recently run a short test integrating Stable Diffusion into their service, whatever their implementation of it was, it worked very well, but it's offline for tweaks right now.
Any updates to get it to work kn amd?
What does V scale do?
Try changing it! I noticed that higher and higher values made the result look worse (artefacts).
it's supposed to make the result more similar to the text prompt but when I try it, the higher value just makes the image look cursed
Working perfectly on my RTX 3060(12GB) at about 3.47it/s, but with a hard limit at 512 of resolution and 250 steps. It's awesome nonetheless. Thank you for your quick and awesome work.
Edit: I've been able to get 704x704 images but they go at 1.15s/it. It's a lot slower, not worth it for me.
Did you try turning on the "Use Half-Precision" ? This makes it possible to be used in hardware that is less than 10GB of VRAM, maybe with this enabled you can achieve grater resolutions and/or speeds because it will have less precision, so less vram is needed. (enabling this will do float16 instead of float32).
If you want more speed, I guess, use the same resolutions with half-precision enabled... I cant test this myself, because I only have 8GB VRAM, and GUI dont let me uncheck it
Does it work on a GTX 1650? Because it only lets my do 62 x 62 and under and even then it just makes a black screen. Does CUDA have to be installed to make this work because its giving me a CUDA memory error when I'm running it as 512 x 512, and I don't know why.
same i NEED ANSWERS
because original stable-diffusion need 10GB+ VRAM, and this one with some optimisations need 6+ too on 512x512
I finally got it to work on my 1650 using this link. 512 is still the limit, but it's finally working
Can you make it so it supports over clocked gpu
What happens if your gpu is overlocked?
It doesn't allow you to render images, instead the image is just black.
Quality submission but the game jam is over... jokes aside...
Thank you for sharing this fantastic Graphic User Interface!
Is there any way to use this without needing an Nvidia driver? Every time I try to generate an image it tells me: RuntimeError: Found no NVIDIA driver on your system. Please check that you have an NVIDIA GPU and installed a driver. If not then that's fine.
this program only works with Nvidia graphics cards so that could be the problem, if you have an Nvidia gpu you can try to install the latest drivers with geforce experience and that might fix it.
thank you this is legendry i hope my graphic card run it
Works perfectly on a GTX 1080.Thanks for creating & sharing this.
does it support img2img?
When I generate an image at 150 steps with a resolution of 512X512 it only shows a black screen. I don't know if it's my GPU (RTX 2060) but it won't work.
Well, It seems that I have found a solution to my own problem
Apparently, it doesn't like it when it has an OVERCLOCKED GPU.
Thanks so much for this! Is there a config where I can disable nsfw content from generating?
just dont use words in your prompt that would make nsfw happen, other then that i dont think thats an option.
All of the images i am trying to render keep on turning out as a fully black square. My GPU is a GTX 1650, is there any fix for this?
Use google collab or try to install it without the GUI. This GUI is good but the default and only option of "half precision" makes it impossible to generate images in some nvidia graphic cards (GTX 1650, 1660, including SUPER) it happened to me, the solution was installing the code locally without any GUI, there are some websites that explain it:
How to Run Stable Diffusion Locally to Generate Images (assemblyai.com)
or you can watch some youtube videos too .
Once you have it running, in the terminal use the command --precision full. That should solve the issue it did it for me! good luck out there and wait for the new version of this GUI this is just the 0.1v
Would it be possible to add an option to link the result to txt2imghd ( https://github.com/jquesnelle/txt2imghd ) for upscaling ?
Also are the text input compatible with prompts from https://promptomania.com/stable-diffusion-prompt-builder/ ?
The promptomania prompts work perfectly good
Cheers for this, works perfectly! Looking forward to img2img implementation
Legendary, thank you!
Hi, do you think you could implement the optimized version?
Hi, I'm getting black images but my card is a 1650 instead of a 1660.
same
same
same
I have a 1650 and my images were showing up green. Using the "--precision full" argument enabled the output, but I'm not sure how you could do that with this build
Will full precision option be added soon?
what's v scale?
CFG scale
I’m not much for legal jargon. I’ve read the full license, when you state “You may re-distribute the weights and use the model commercially and/or as a service. If you do, please be aware you have to include the same use restrictions as the ones in the license and share a copy of the CreativeML OpenRAIL-M to all your users (please read the license entirely and carefully)”.
Does this pertain to the produced images being sold or the AI model itself being used by users of the GUI such as myself or others? Like do I need to put a link in my descriptions for craft supplies that this was used and put the copy of the license in there?
This license is a copy from the Stable Diffusion team, this line is just about the AI model.
Images generated by the model are explained on the second line of the license. You can use the generated images as you like, no credit needed, but if you have more questions about the license, you need to talk with the Stable Diffusion team.
Thank you so much! I’m really excited for what this does for my use case. Looking forward to where it goes from here and I appreciate all of the work done to make a GUI and contained executable.
This is one of the coolest shit i seen, i been just playing around all night trying to get cool results, thanks a lot for soing this! i been trying to follow the guides from github but my non-programer brain has been strugling to get it to work, this works awesomelly!
Glad you are enjoying it.
Nice work, hopefuly future models / algorithm will allow for better output resolution thant 512x512 on a RTX 2070 Super. 8GB seems to be a little short for that ? (that's kinda crazy)
Actually is a miracle a model like that can run in a consumer card, even so its probably will improve a little the memory usage in the future.
Hello! How can I specify the number of copies to be generated?
For now only one copy per prompt, you can repeat the prompt in multiple lines if you want.
Why does it render 2 samples, one is mine and the other is different from what I wrote? as if it was a random one or written by someone else.
I have the same issue - happened after I chose to generate 4 ideas at a time. After that if I switch to just 1 it still tries to generate 4 ideas total where the first one is what you asked for and the rest is random stuff unrelated to the promp
Did you manage to fix it?
The problem is you put an empty line after the prompt. It basically takes an empty prompt and generates another image
Perhaps there is a empty line in your prompts window?
I had the same issue, I realized I had an extra line without text so it would generate my initial prompt then move onto the second line, since theres no text it would just generate an incoherent image.
Are you releasing higher resolution (more than 512x512) in the next updates / versions?
Thank you by the way!
You can already use it, if your card support it.
There's no need. What people usually do is use a different neural network that up-scales images like Upscale Pics or UpscaleMedia
Is the 1660 series issue something that should be fixable in a future update, or is it a hardware issue with the card?
Its a hardware problem with half precision I think. The fix is to use full precision model, but it will take even more memory to use the app.
Ok thanks! Are you anticipating enabling full precision in a future update? If that's technically possible, I'd love to have the option to run Stable Diffusion through this program even if it takes a long time because of the amount of memory required.
Great work! Very fast on my 2080super. Thank you!
Every time I download it says the .rar is damaged or corrupt. Any idea why? Update: my program WinRar was out of date. My bad!
Could you please let me know if there is a way to use cli mode?
Not for now, probably in the future
AMAZING! WORKED WITH 0 ISSUES!
I get 1.02 it/s with my GTX 1060, what will it be like with a 3060?
also, what's V-scale?
with laptop 3060 got ~2.8 it/s on steps 100 512x512
Oh, nice, I need a new computer
My desktop 3060 gets 3.58 it/s at 512x512.
Arround 3.47it/s with my 3060 in desktop.
On my 3090Ti at 512x512 I get ~6.5 it/s.
3070 I get 4.2 ish.
I signed up here specifically to thank you! It works easy with my 3090, i love that the promt text stays after generating, very easy to work on the text !
Would be good if you could simply click on a few different resolutions, instead re-entering manually.
Looking forward to the next update! :D
Improvements for the resolutions will be available in a few updates, thanks for the compliment =)
I expected you would be one of the first to make a GUI for this, good job! A great feature for a future update would be the ability to import an image and perform the AI transformations on that, some other GUIs have that feature so I imagine it is possible to implement
I didn't think anyone would think of me for this, thanks.
Yeah, img2img will be available in a few updates.
Yes please! Image to Image would be amazing, that's my top requested feature.
This GUI app is amazing, thank you so much for putting this together!