3.5x speed upgrade with Nvidia .dll replacement working for InvokeAI Standalone 3.2.0 and a 2x speed increase for standalone 2.3.2 and the regular install of 2.3.4.
I've replaced the CUDA files for invokeAI 2.3.0 and it makes it 3.5x faster with my Nvidia 4090. I'm running InvokeAI 2.3.0 at 28 it/s just by replacing the CUDA .dll files alone. However, I cant seem to get the latest to be as fast by replacing the .dll files but at least it's still 2x faster on InvokeAI 2.3.2 and also the normal install of the latest 2.3.4 with .dll replacement. Not sure why I cant get it going as fast as 2.3.0 at 3.5x the speed. For in-painting I'm still using Invoke 2.3.0 with the upgraded .dll's and loving the 3.5x speed. Any idea why the versions after 2.3.0 are not performing as fast as InvokeAI 2.3.0 with the latest CUDA files? anyone know how to get more speed out of Invoke 2.3.2-2.3.4?
I Downloaded the Nvidia zip file here and used the .dll files found in the bin folder:
I figured it out. The improvement you are able to see is because nvidia fixed the driver compatibility issue with CUDA 11.8 on the 4k series of cards. Everyone using a previous generation is already running at maximum performance.
So far my experience has been amazing. Very user friendly and fun to use! Are the use of LoRAs or Dreambooth style training something you are considering?
You can run the update_experimental.bat, then you get Lora support. :) Or, if you installed via the itch app, you can also wait some days, then I get the patch ready.
I have it installed through the itch app. Would it mess things up if I update through the .bat file? Approximately when do you think you'd have it updated on itch?
The release has been updated. Make sure you run "update_experimental.bat" from your invokeai folder.
Chose option 1 and press enter. The update process will begin.After not too long your will receive a message informing you the update was successful.
You may now launch InvokeAi using the launcher (do not pay attention if the launcher says version 2.3.2, it's just a bug). What matters is the version displayed in the Web UI, in the upper left corner.Do not forget to post your reviews, questions and your own creations here.
For models please visit Civitai. Feel free to start using InvokeAI with a powerful model:
did the errors I encounter are corrected now? :3 , should try to reinstall? I dont care doing a few steps before start generate images tho. Im glad this works cz wasn't able to install via github but whatever xD
I was able to update via -update_experimental.bat- now, a few days ago this wasn't posible, when u try to start after installation, on the console it will appear an error, ("tuple index out of range") and ask u if u want to (re)install the models, just type n for cancel that, it will open the page, however, there's a little "bug" with the token "thing" i think, u have to load a ckpt using model manager, then click on update model, that solve the issue, then change the sampler option, lets say ur using dpm++ sde karras, change for whatever option, generate an image and then go back to the one you use, if not the image will be generate really bad xD, sorry if this are broken english, isn't my main languaje :3
Hello, thank you very much for your comment, I am oriented to update this program, I had been looking for days how to do it, also for the clarifications apart to correct it, by the way my original language is not English either so we are the same XD.
sometimes this works wonderfully for me, while others i can't seem to get it to stop generating empty black images and i wish i understood why. using a 2080 super. thanks!
Sometimes it depends on the checkpoints you have installed. I recommend updating to version 2.3.4 since it is also faster. You should convert checkpoint to diffusers for maximum speed, it literally halves the rendering time.
i Use Itch to install Invoke AI just because it says i can easily upgrade InvokeAI. but now 2.3.4 is out, when I click "update" here , nothing happened. it said , nothing new. ...
That's because this is a standalone version "by Sunija". So I guess he will update it when he can. You can download the other one as well, I ran both for a while.
You can update with the "update_experimental.bat" in the standalone folder. :) But I'll also upload the update tomorrow (and integrate a button for the update script).
I am newbie, i use Itch just because it says we can update InvokeAI as easy as like upgrade a game (when new version is out) , however now like you said 2.3.4 is out, but I click "update" button here, nothing new.
or maybe wait a couple of days? or something I did wrong? thank you.
The current version of mine installed via Itch is 2.3.2 ...
Hm... that looks all fine. You could try to reinstall it (make a backup of your output folder first!) or poke me on Discord (Sunija#6598) and we solve it together. :)
It cannot run on phones, because they don't have the required performance. :( You can run it on your PC and then access it on your phone via the link in the starter.
1) Run the starter as you normally would. 2) Check the "Share access" option on the left side. 3) Restart the starter. 4) Open the link next to "Your Wi-Fi:" on your phone.
Sadly, I don't know. :( I hoped they would be available months ago. Apparently there are some less fancy UIs that already support AMD cards. You could google about those (or somebody posts them as an answer here).
File "E:\Itch\Invoke_Ai\invokeai\invokeai_2_3_0_standalone\env\lib\site-packages\invokeai\backend\modules\parameters.py", line 45, in parameters_to_command
if "init_img" in params and len(params["init_img"]) > 0:
TypeError: object of type 'Image' has no len()"
Suffice to say I have put in an image into the prompt and also tested various settings, result is the same seemingly regardless of model used.
Of note, Unified canvas still works, so it can still generate based on another image, just something going wrong with img 2 img specificly.
I read this line at the beginning of the threat: "Requires a Nvidia Graphics Card (except GTX 1650)." what's this mean, does it mean that GTX 1650 is not gonna work lmao. I use 1650 lol:"))) please answer me thanks
It won't work. :( The 16xx generation has a bug that makes it need more VRAM than other cards. The 1650 is the only model that doesn't have enough VRAM to do that. :(
So, today marks it one week since i downloaded this and I got to say, i only slept 12 hours in one week. I've been stuck to my computer ever since. After 3 days of using the cpu to render a picture in 55minutes, I went and I bought a 3060 without even looking at the money(lucky the card was 150$ off on sale). My only regret is that they didn't have a 3070 in stock. I just want things to load faster!! I'm considering of adding another gpu, maybe even an older one as long as it runs the latest cuda drivers.
One question: i would like to compile my own sampler, where could i find the source code for the existing samplers ?
Not to forget, a huge THANK YOU bro, for making this software real.
Cool that you like it! :) I don't know where to find the samplers in the source code, but I bet the InvokeAI team can help you on their discord. :) Also, they are working on the big 3.0 update that will allow plugins/nodes, so adding a new sampler should be pretty easy by then. :)
And for AI the 3060 is usually better than the 3070. :) The 3060 has 12 GB of VRAM (compared to the 8 GB of the 3070), which allows for *much* bigger images than the 3070 would. :)
Yo I found out about this project yesterday and used it a couple hours already, pretty good work, btw I already read a comment where you answered that the Loras support will be added on the next big release which was my only question so nice job overall, really looking forward to this and is it okay to try to get involved in this somehow? I'm a software engineering student and this really caught my attention, best wishes!! Hope to see more of this :)
It's a great launcher but i do feel like it's mainly for beginners in ai art because as someone who is familiar with webui1111 there are some few things i dislike/like about it.
What i don't like about it:
1) No ability to read metadata on an image that wasn't generated inside the program (couldn't find the tab to do so anywhere)
2) Low quality upscalers , yes there are the options to 2x and 4x an image but the upscaler just isn't good and a good upscaler is a lot more important than the initial generated image. (Nearest, latent, 4xultra, animesharp.. all missing)
3) Loras, can't load loras
4) Ability to refresh embeddings once the program is already working
5) Sampler names are all shortened which i don't understand how anyone is supposed to know what is k_lms/k_huen/etc...
What i do like about it:
1) Generates images really fast (4s on a 512x512 image on a gtx 2070 super)
2) Unified canvas (amazing feature)
3) The UI in general is very comfy looking and inviting, it feels like having your own private Novel AI
4) Lots of tabs to help with image generation
5) You can load the model plus the chosen vae of that model without having to do so manually
In summary for beginners this is a 9/10 straight out great but for people who are accustomed to webui1111 high quality upscaling and lora usage this is a 6/10.
Can't wait to see more, and I'm sure one day it will be a 10/10! :)
Thanks for the detailed review! :) I'll let the InvokeAI team know about it. LORAs will come with the next big upgrade, but e.g. I didn't think about the upscalers, so your comment definitely helped! :)
hey so i love what you did here it makes this an easy and accessible way to use AI art but i was wondering if there was a way to change my default model i like stable diffusion but i use anythingv4.5 more if you can please help me out i would appreciate it
Soooooo, first thank you sincerely for both InvokeAI and Aiimages :)
Having some tech-diffs, to put it mildly - actually a bit of a pre melt-down panic level event is fast brewing! I have been using InvokeAI v4 for a few weeks, not updated to v5 yet, and tonight I opened up all 3 instances of Aiimages (using different models per instance to save on hassle of model switching) that had in the region of 60k+ generations between them. Intead of my vast wonderful galleries which were all there yesterday, I was met with 3 completely blank galleries and each one populated with the orig InvokeAi prompt - Content: cute owl, Style: watercolor painting.
PLEASE tell me that someone else has previously encountered this issue and there's a way to retrieve the galleries. The generated images are still there in the Output folders and both dream_log.md and dream_log.txt are still intact and contain all of my prompts for all 3 Aiimages instances I've been working in. Anyone got a fix for this? Otherwise I've just lost a vital chunk of half a years work.....
Kind dev, please let me know if this catastrophe can be fixed somehow. <3
InvokeAI or Aiimages? :X First the good news: As long as the images are still there, nothing is lost on the InvokeAI site. Just to be sure: Make a backup before you continue. :3
You can set an output folder in the bottom left of the starter. Set this to the folder that InvokeAI should use (= the one where your outputs are).
If you really meant Aiimages, then we can look at that again. :) If you need more help just poke me on Discord (Sunija#6598) or write here. :)
Sorry, I appreciate why that might have been confusing, I was a little panic-stricken as I was typing last night!
Are you aware of the disappearing history issue in Aiimages (possibly caused by a crash or by opening Aiimages and InvokeAI at the same time), and have you or any of the users on Itch or Discord found a method that would make it possible to restore my history in each of the Aiimages instances?
I think I may also have accidentally opened both InvokeAi and then Aiimages at the same time last night, which I've been avoiding, as from past experience, I know there's a shared file hiding somewhere outside the Aiimages folders, and have seen prompts and settings being populated between each instance of Aiimages when switching the instance of Aiimages I'm working in. I suspect that may be the issue, as upon opening Aiimages presently, the initial settings for InvokeAi appear, with the "cute owl" prompt loaded, and all the images generated are no longer there in history.
Thank you for reading my lengthy reply(s)! Hope my situation is more clear
Post Coffee Edit: Separated feedback and observation for both programs in to a separate post below :)
Currently I'm using both of your standalone offline tools. I have been generating material in Aiimages over the last half a year, and this month have started using InvokeAi mainly for the Unified Canvas feature. Taking images made in Aiimages, and importing them to InvokeAI to develop them further with the canvas toolset.
I have three instances/installs of Aiimages that I use for separate content/workflows using different models, and also because after 20k+ images are in history, in my experience it gets a little slow and frequently crashy when scrolling through the vast galleries.
With InvokeAi, I find its design more of a hassle for accessing earlier generations because it only loads 50 (or is it 100?) images in to cache, and then I have to constantly scroll down and click to load another 50/100 images. I generate LOTS of images per prompt so accessing the images becomes cumbersome in InvokeAi. Very klunky and limiting in contrast with Aiimages, in that respect. Pros and Cons yo!
For my workflow, being able to load up one of the Aiimages instances and quickly scrolling the history to get the precise prompt for an image to work on inside InvokeAi's unified canvas has been invaluable. So while I still have the images, and still have the dream_md / .txt prompt histories for each of the three Aiimages instances intact in their respective folders, I can no longer easily scroll the history to access the image with corresponding prompts.
Aiimages handles the Style/Content fields a little differently to InvokeAi with its Prompt/Negative fields. For instance, if I copy a Style/Content prompt from Aiimages into InvokeAI's prompt field, all is well until I reload the prompt at a later stage, as it then takes the original prompting character/token sequence and populates the Negative prompt field with any [negatives] from the original Aiimages prompt. It appears to modify the generated output aesthetic quite dramatically at times when this happens, so I usually just copy the prompt from Aiimages back in to the InvokeAi prompt field each time I want to return to a theme (say for going past the 1024 x 1024 limit in Aiimages) to keep the precise prompt.
If I'm not mistaken, this also means the metadata when taking an image from one application to the other changes the numerical(/character?) string that the prompts are converted into, so I very much hope to retain the history of images generated in my Aiimages instances over the last 6 months inside Aiimages for easy access of/reference to the original prompts.
Thanks so much for releasing these tools for the beginners, the Unity devs in the case of Aiimages, and for the folk who just like to run their machines offline to avoid the overlords snooping and auto-update tampering with their delicate work environments! :D x
First the bad news: There is no easy way to restore the aiimages history file. :X The good news: All those images still have their prompts/settings in the meta info. And InvokeAI can read those meta infos.
So I guess the best way atm to reproduce your workflow would be: 1) Open your aiimages output folder in your Windows Explorer, set the images to maximum size. 2) If you want to use the prompt of an image, copy it into the invokeai output folder. 3) In InvokeAI, you can click on the image and use the "use prompt" button on top.
Let me know if that workflow works for you. :X
In theory it could be possible to scan a folder, read the metadata and rewrite it into an aiimages history file. But that would be quite a bit of work for a deprecated tool. :X
As a temp workaround that could well be the solution, good thinking and thanks for the suggestion!
Sad/bad news indeed on all fronts, especially re: Aiimages' deprecated status D:
I absolutely love the design of Aiimages, and as a Unity dev I'm still planning to integrate it in to some of my projects. In the Unity space it's still IMHO the best SD tool available, by far, and I've tried them all (except Keijiro Takahashi's recent addition for the Mac folk). Props are due for the work you've put into it!
My waifu is fairly experienced with Python. If you have a sense of what would be involved to scan and rewrite the history file, would you be ok with sharing your pseudo-code thoughts on what steps would be required in Python to achieve the history restoration?
If we have any luck achieving it, we'd of course share the scripts with the community in case anyone else encounters the issue.
Hi! When I try to load the image I get this message:
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 16.00 MiB (GPU 0; 6.00 GiB total capacity; 4.34 GiB already allocated; 13.56 MiB free; 4.41 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
>> Could not generate image.
>> Usage stats:
>> 0 image(s) generated in 3.99s
>> Max VRAM used for this generation: 4.66G. Current VRAM utilization: 4.66G
>> Max VRAM used since script start: 4.66G
The image doesn't load and it stays grey :< Is there anything I can do? I have GeForce GTX 1660 Ti
Sounds like you're low on VRAM. :3 There are several things that you can try:
1) Close InvokeAI and check how much VRAM is already used up by the system in your task manager. Try to close stuff that you don't need and check if it lowers the VRAM usage.
2) Disable the NSFW checker, as it also needs some VRAM. You have to restart InvokeAI so the setting is applied.
3) Disable the in-progress preview, because it also needs VRAM. You can disable it in the tool on the cogwheel in the top right -> Display In-Progress Images -> None
4) There is a --free_gpu_mem option that will save VRAM. It will slow the generation a bit, though. To activate it, you have to open invokeai_2_3_0_standalone\invokeai\invokeai.init in a text editor and add the line after one of the upper rows. But I'd first check if the rest works. :3
This can be (random) cause when you set image size out of range. Try setting the image size between these number (512, 768, 1024). When image size out of these number sometime it work, sometime it won't. Make the prompt longer also fix this problem. And finally if it doesn't fix, your VRAM might already full.
Just a side note, after reading about Windows antivirus below. Bitdefender saw two of the Python files as Ransomeware. I added then them as OK and the install continued as normal.
HI. Unfortunately, when I launch Invoke AI I get an error, specifically PermissioneError WinError 10013 Attempt to access the socket in ways not allowed by the respective access permissions. I don't know what to do unfortunately
Thanks for availability, my discord is ElPolloLoco#0042, I tried to run it as administrator, disable windows defender and windows firewall but it still doesn't work
I'm having a problem with the easynegative embedding and as the only thing I found about this is another comment, i'm trying my luck too: >> Notice: embeddings/easynegative.pt was trained on a model with an incompatible token dimension: 768 vs 8 The embedding does not show in the list and does not seem to do anything when I try to use it anyway. (No ">> Preparing tokens for textual inversion")
Guys, if your trying to use an AMD gpu it will not work on windows (for now). But if you want to use stable diffusion i cant recommend Google colab (Stable diffusion) enough.
← Return to tool
Comments
Log in with itch.io to leave a comment.
AMD announced ROCm will be coming to windows. Will support be added?
I think InvokeAI will definitely add it once ROCm properly works on windows. :)
3.5x speed upgrade with Nvidia .dll replacement working for InvokeAI Standalone 3.2.0 and a 2x speed increase for standalone 2.3.2 and the regular install of 2.3.4.
I've replaced the CUDA files for invokeAI 2.3.0 and it makes it 3.5x faster with my Nvidia 4090. I'm running InvokeAI 2.3.0 at 28 it/s just by replacing the CUDA .dll files alone. However, I cant seem to get the latest to be as fast by replacing the .dll files but at least it's still 2x faster on InvokeAI 2.3.2 and also the normal install of the latest 2.3.4 with .dll replacement. Not sure why I cant get it going as fast as 2.3.0 at 3.5x the speed. For in-painting I'm still using Invoke 2.3.0 with the upgraded .dll's and loving the 3.5x speed. Any idea why the versions after 2.3.0 are not performing as fast as InvokeAI 2.3.0 with the latest CUDA files? anyone know how to get more speed out of Invoke 2.3.2-2.3.4?
I Downloaded the Nvidia zip file here and used the .dll files found in the bin folder:
https://developer.download.nvidia.com/compute/redist/cudnn/v8.7.0/local_installers/11.8/
Then I replace the old .dll files with the new ones into the following: InvokeAI\venv\Lib\site-packages\torch\lib
I'm testing this for sure. Where did you find version 2.3.0
The improvement i'm seeing is by 1.5 seconds. Sometimes slower depending on the CUDA dlls version.
I figured it out. The improvement you are able to see is because nvidia fixed the driver compatibility issue with CUDA 11.8 on the 4k series of cards. Everyone using a previous generation is already running at maximum performance.
So far my experience has been amazing. Very user friendly and fun to use! Are the use of LoRAs or Dreambooth style training something you are considering?
You can run the update_experimental.bat, then you get Lora support. :) Or, if you installed via the itch app, you can also wait some days, then I get the patch ready.
Amazing work! I'll wait for the update on Itch. :D
I have it installed through the itch app. Would it mess things up if I update through the .bat file? Approximately when do you think you'd have it updated on itch?
Sweet, any eta? Wishing you the best :)
How can I install the patch via itchi app? is there a button? Help plz!
You can double-click update_experimental.bat in the invokeai folder. :)
its not working and I did all of the video, takes like half hour and still no 1 pic, I have to sleep.
I'm not sure I understand. What's not working ? The software or the update ? Join me on discord, I'd be glad to help.
The release has been updated. Make sure you run "update_experimental.bat" from your invokeai folder.
Chose option 1 and press enter. The update process will begin.After not too long your will receive a message informing you the update was successful.
You may now launch InvokeAi using the launcher (do not pay attention if the launcher says version 2.3.2, it's just a bug). What matters is the version displayed in the Web UI, in the upper left corner.Do not forget to post your reviews, questions and your own creations here.
For models please visit Civitai. Feel free to start using InvokeAI with a powerful model:
Crystal-Clear_v1 | Stable Diffusion Checkpoint | Civitai
did the errors I encounter are corrected now? :3 , should try to reinstall? I dont care doing a few steps before start generate images tho. Im glad this works cz wasn't able to install via github but whatever xD
I was able to update via -update_experimental.bat- now, a few days ago this wasn't posible, when u try to start after installation, on the console it will appear an error, ("tuple index out of range") and ask u if u want to (re)install the models, just type n for cancel that, it will open the page, however, there's a little "bug" with the token "thing" i think, u have to load a ckpt using model manager, then click on update model, that solve the issue, then change the sampler option, lets say ur using dpm++ sde karras, change for whatever option, generate an image and then go back to the one you use, if not the image will be generate really bad xD, sorry if this are broken english, isn't my main languaje :3
Hello, thank you very much for your comment, I am oriented to update this program, I had been looking for days how to do it, also for the clarifications apart to correct it, by the way my original language is not English either so we are the same XD.
sometimes this works wonderfully for me, while others i can't seem to get it to stop generating empty black images and i wish i understood why. using a 2080 super. thanks!
Sometimes it depends on the checkpoints you have installed. I recommend updating to version 2.3.4 since it is also faster. You should convert checkpoint to diffusers for maximum speed, it literally halves the rendering time.
i Use Itch to install Invoke AI just because it says i can easily upgrade InvokeAI. but now 2.3.4 is out, when I click "update" here , nothing happened. it said , nothing new. ...
That's because this is a standalone version "by Sunija". So I guess he will update it when he can. You can download the other one as well, I ran both for a while.
tHANKS
You can update with the "update_experimental.bat" in the standalone folder. :) But I'll also upload the update tomorrow (and integrate a button for the update script).
Update 2.3.4 is out !! Now with LORA and LYCORIS support.
Get it here: Release InvokeAI Version 2.3.4 - A Stable Diffusion Toolkit · invoke-ai/InvokeAI (github.com)
Thanks WarAnakin,
I am newbie, i use Itch just because it says we can update InvokeAI as easy as like upgrade a game (when new version is out) , however now like you said 2.3.4 is out, but I click "update" button here, nothing new.
or maybe wait a couple of days? or something I did wrong? thank you.
The current version of mine installed via Itch is 2.3.2 ...
Hey, great tool, by im curious is it possible to generate image with transparent background? Like UI elements, icons etc..
I don't think so. But you can try to generate something with an even background and then use another ai to cut out the foreground. :)
Any idea when the launcher will be updated via itch?
Tomorrow, I hope. :)
But you can already update today by running the update_experimental.bat in the invokeai_standalone folder.
Any chance the ControlNet or OpenPose add-ons?
In the future, yes. :) Version 3.0 will support plugins, which should allow stuff like ControlNet.
I was using perfectly, but stopped works and now keepping saying "System config requested"
What should i do?
Can you send me a screenshot of the error on Discord (Sunija#6598)? :3
unified canvas does not works anymore
Hm... that looks all fine.
You could try to reinstall it (make a backup of your output folder first!) or poke me on Discord (Sunija#6598) and we solve it together. :)
what is DPM++ SDE Karras and which sampler in the list is it?
I'm confused of the names
it's called k_dpmpp_2
"k" means Karras and "pp" means plus plus
Version 2.3.3 is out, change notes: Release InvokeAI Version 2.3.3 - A Stable Diffusion Toolkit · invoke-ai/InvokeAI (github.com)
Its not updating via itch. Any idea why?
Probably Sunija is working on his launcher.
Will there be an apk version
It cannot run on phones, because they don't have the required performance. :(
You can run it on your PC and then access it on your phone via the link in the starter.
can I go into more detail?
1) Run the starter as you normally would.
2) Check the "Share access" option on the left side.
3) Restart the starter.
4) Open the link next to "Your Wi-Fi:" on your phone.
Does that explain it? :)
Hey, is there any API to integrate it in your own code
The next big upgrade will bring a node system that should make it really easy to add your own code (as nodes).
hey, there is any prevision of when the support for amd cards gonna be up ? ;-;
Sadly, I don't know. :(
I hoped they would be available months ago. Apparently there are some less fancy UIs that already support AMD cards. You could google about those (or somebody posts them as an answer here).
Hi, img 2 img doesn't work, always get an error up by the model, and it fails. Command prompt has this at the end each time.
"command = parameters_to_command(generation_parameters)
File "E:\Itch\Invoke_Ai\invokeai\invokeai_2_3_0_standalone\env\lib\site-packages\invokeai\backend\modules\parameters.py", line 45, in parameters_to_command
if "init_img" in params and len(params["init_img"]) > 0:
TypeError: object of type 'Image' has no len()"
Suffice to say I have put in an image into the prompt and also tested various settings, result is the same seemingly regardless of model used.
Of note, Unified canvas still works, so it can still generate based on another image, just something going wrong with img 2 img specificly.
Any chance the ControlNet or OpenPose add-ons?
Why I faced "An exception has occurred: 'cp950' codec can't decode byte 0xe6 in position 474: illegal multibyte sequence" problem ?
Is there any solutions?
Can you send me a screenshot of the error on Discord (Sunija#6598)? :)
Hi,
I tried some images and they all came out blurred with a big exclamation mark.
what do I do?
turn off nsfw filter.
thanks!
Are you ever going to add support for AMD gpus? It really sucks not being able to use this on my 6600XT. 😭
I am also waiting for AMD support to use my 3 6700XT.
For now, I am using the Stable Diffusion GUI which has support
Any idea where I can download that?
Just search right here on itch for "Stable Diffusion GUI"
I recommend taking a look at their Discord which has a lot of content for those just starting out with it
i have installed it but what should i do now:/ - can someone help me i am to dumb for that😂👍
This video should explain everything. :)
I read this line at the beginning of the threat: "Requires a Nvidia Graphics Card (except GTX 1650)."
what's this mean, does it mean that GTX 1650 is not gonna work lmao. I use 1650 lol:")))
please answer me thanks
It won't work. :(
The 16xx generation has a bug that makes it need more VRAM than other cards. The 1650 is the only model that doesn't have enough VRAM to do that. :(
Worst case scenario, use your cpu, you'll be able to render mostly anything, it will just take a long time...
Not to lead away from this program, but NMKD Stable Diffusion (here on Itch) has settings for low-memory cards as well as AMD cards.
I guess its not really a lead away if they literally cant use this haha.
So, today marks it one week since i downloaded this and I got to say, i only slept 12 hours in one week. I've been stuck to my computer ever since. After 3 days of using the cpu to render a picture in 55minutes, I went and I bought a 3060 without even looking at the money(lucky the card was 150$ off on sale). My only regret is that they didn't have a 3070 in stock. I just want things to load faster!! I'm considering of adding another gpu, maybe even an older one as long as it runs the latest cuda drivers.
One question: i would like to compile my own sampler, where could i find the source code for the existing samplers ?
Not to forget, a huge THANK YOU bro, for making this software real.
Cool that you like it! :)
I don't know where to find the samplers in the source code, but I bet the InvokeAI team can help you on their discord. :) Also, they are working on the big 3.0 update that will allow plugins/nodes, so adding a new sampler should be pretty easy by then. :)
And for AI the 3060 is usually better than the 3070. :) The 3060 has 12 GB of VRAM (compared to the 8 GB of the 3070), which allows for *much* bigger images than the 3070 would. :)
Do you happen to know if i add an extra card, will it help with the rendering time ? or it will just render a 2nd image on the side ?
Yo I found out about this project yesterday and used it a couple hours already, pretty good work, btw I already read a comment where you answered that the Loras support will be added on the next big release which was my only question so nice job overall, really looking forward to this and is it okay to try to get involved in this somehow? I'm a software engineering student and this really caught my attention, best wishes!! Hope to see more of this :)
Thanks!
You can join the InvokeAI discord, I bet they are always grateful for help! <3
Hey my windows defender has been detecting trojan wacatac b ml on your requeriments_checker, what is this?
Thanks for the report! :)
Can you send me a screenshot of the report on Discord (Sunija#6598)? My windows defender doesn't find anything there, so I cannot check myself. :(
And are you sure that it's the requirements checker? :3 There is a known issue that some of the AI models are falsely flagged as threats.
Also, in case somebody is interested, that's the whole script of the requirements checker / starter: https://github.com/GothaB/invokeai_starter/blob/main/invokeai_starter/MainWindow...
i sent you a friend request so i can show you the print
Hi! Will the DPM++ 2M Karras sampler be added?
Already in there! :)
It's k_dpmpp_2.
Latest version gives me a "ImportError: DLL load failed while importing cv2" and won't start
Thanks for reaching out! :)
Can you send me a screenshot of the error on Discord (Sunija#6598), so we can fix it together? :)
Until then, you can start the AI via the "helper.bat" file next to the starter. :3
It's a great launcher but i do feel like it's mainly for beginners in ai art because as someone who is familiar with webui1111 there are some few things i dislike/like about it.
What i don't like about it:
1) No ability to read metadata on an image that wasn't generated inside the program (couldn't find the tab to do so anywhere)
2) Low quality upscalers , yes there are the options to 2x and 4x an image but the upscaler just isn't good and a good upscaler is a lot more important than the initial generated image. (Nearest, latent, 4xultra, animesharp.. all missing)
3) Loras, can't load loras
4) Ability to refresh embeddings once the program is already working
5) Sampler names are all shortened which i don't understand how anyone is supposed to know what is k_lms/k_huen/etc...
What i do like about it:
1) Generates images really fast (4s on a 512x512 image on a gtx 2070 super)
2) Unified canvas (amazing feature)
3) The UI in general is very comfy looking and inviting, it feels like having your own private Novel AI
4) Lots of tabs to help with image generation
5) You can load the model plus the chosen vae of that model without having to do so manually
In summary for beginners this is a 9/10 straight out great but for people who are accustomed to webui1111 high quality upscaling and lora usage this is a 6/10.
Can't wait to see more, and I'm sure one day it will be a 10/10! :)
Thanks for the detailed review! :)
I'll let the InvokeAI team know about it. LORAs will come with the next big upgrade, but e.g. I didn't think about the upscalers, so your comment definitely helped! :)
hey so i love what you did here it makes this an easy and accessible way to use AI art but i was wondering if there was a way to change my default model i like stable diffusion but i use anythingv4.5 more if you can please help me out i would appreciate it
Top right in the InvokeAi webui, there's an isometric cube icon for "Model Manager".
Click that, you'll see "+ Add New".
Right click in the relevant Windows folder path containing your model(s), and select "Copy Address".
Ctrl-V to paste into the "Checkpoint Folder" field and hit Enter.
You should now be able to add your model of choice, and all going well, it should then appear in the drop down, to the left of the Model Manager icon.
Hope that helps ;)
Soooooo, first thank you sincerely for both InvokeAI and Aiimages :)
Having some tech-diffs, to put it mildly - actually a bit of a pre melt-down panic level event is fast brewing! I have been using InvokeAI v4 for a few weeks, not updated to v5 yet, and tonight I opened up all 3 instances of Aiimages (using different models per instance to save on hassle of model switching) that had in the region of 60k+ generations between them. Intead of my vast wonderful galleries which were all there yesterday, I was met with 3 completely blank galleries and each one populated with the orig InvokeAi prompt - Content: cute owl, Style: watercolor painting.
PLEASE tell me that someone else has previously encountered this issue and there's a way to retrieve the galleries. The generated images are still there in the Output folders and both dream_log.md and dream_log.txt are still intact and contain all of my prompts for all 3 Aiimages instances I've been working in. Anyone got a fix for this? Otherwise I've just lost a vital chunk of half a years work.....
Kind dev, please let me know if this catastrophe can be fixed somehow. <3
Thanks for reaching out!
InvokeAI or Aiimages? :X First the good news: As long as the images are still there, nothing is lost on the InvokeAI site. Just to be sure: Make a backup before you continue. :3
You can set an output folder in the bottom left of the starter. Set this to the folder that InvokeAI should use (= the one where your outputs are).
If you really meant Aiimages, then we can look at that again. :) If you need more help just poke me on Discord (Sunija#6598) or write here. :)
Thanks for the reply!
Sorry, I appreciate why that might have been confusing, I was a little panic-stricken as I was typing last night!
Are you aware of the disappearing history issue in Aiimages (possibly caused by a crash or by opening Aiimages and InvokeAI at the same time), and have you or any of the users on Itch or Discord found a method that would make it possible to restore my history in each of the Aiimages instances?
I think I may also have accidentally opened both InvokeAi and then Aiimages at the same time last night, which I've been avoiding, as from past experience, I know there's a shared file hiding somewhere outside the Aiimages folders, and have seen prompts and settings being populated between each instance of Aiimages when switching the instance of Aiimages I'm working in. I suspect that may be the issue, as upon opening Aiimages presently, the initial settings for InvokeAi appear, with the "cute owl" prompt loaded, and all the images generated are no longer there in history.
Thank you for reading my lengthy reply(s)! Hope my situation is more clear
Post Coffee Edit: Separated feedback and observation for both programs in to a separate post below :)
Currently I'm using both of your standalone offline tools. I have been generating material in Aiimages over the last half a year, and this month have started using InvokeAi mainly for the Unified Canvas feature. Taking images made in Aiimages, and importing them to InvokeAI to develop them further with the canvas toolset.
I have three instances/installs of Aiimages that I use for separate content/workflows using different models, and also because after 20k+ images are in history, in my experience it gets a little slow and frequently crashy when scrolling through the vast galleries.
With InvokeAi, I find its design more of a hassle for accessing earlier generations because it only loads 50 (or is it 100?) images in to cache, and then I have to constantly scroll down and click to load another 50/100 images. I generate LOTS of images per prompt so accessing the images becomes cumbersome in InvokeAi. Very klunky and limiting in contrast with Aiimages, in that respect. Pros and Cons yo!
For my workflow, being able to load up one of the Aiimages instances and quickly scrolling the history to get the precise prompt for an image to work on inside InvokeAi's unified canvas has been invaluable. So while I still have the images, and still have the dream_md / .txt prompt histories for each of the three Aiimages instances intact in their respective folders, I can no longer easily scroll the history to access the image with corresponding prompts.
Aiimages handles the Style/Content fields a little differently to InvokeAi with its Prompt/Negative fields. For instance, if I copy a Style/Content prompt from Aiimages into InvokeAI's prompt field, all is well until I reload the prompt at a later stage, as it then takes the original prompting character/token sequence and populates the Negative prompt field with any [negatives] from the original Aiimages prompt. It appears to modify the generated output aesthetic quite dramatically at times when this happens, so I usually just copy the prompt from Aiimages back in to the InvokeAi prompt field each time I want to return to a theme (say for going past the 1024 x 1024 limit in Aiimages) to keep the precise prompt.
If I'm not mistaken, this also means the metadata when taking an image from one application to the other changes the numerical(/character?) string that the prompts are converted into, so I very much hope to retain the history of images generated in my Aiimages instances over the last 6 months inside Aiimages for easy access of/reference to the original prompts.
Thanks so much for releasing these tools for the beginners, the Unity devs in the case of Aiimages, and for the folk who just like to run their machines offline to avoid the overlords snooping and auto-update tampering with their delicate work environments! :D x
Thanks for the detailed review! :)
First the bad news: There is no easy way to restore the aiimages history file. :X
The good news: All those images still have their prompts/settings in the meta info. And InvokeAI can read those meta infos.
So I guess the best way atm to reproduce your workflow would be:
1) Open your aiimages output folder in your Windows Explorer, set the images to maximum size.
2) If you want to use the prompt of an image, copy it into the invokeai output folder.
3) In InvokeAI, you can click on the image and use the "use prompt" button on top.
Let me know if that workflow works for you. :X
In theory it could be possible to scan a folder, read the metadata and rewrite it into an aiimages history file. But that would be quite a bit of work for a deprecated tool. :X
Hey thanks so much for your reply. :)
As a temp workaround that could well be the solution, good thinking and thanks for the suggestion!
Sad/bad news indeed on all fronts, especially re: Aiimages' deprecated status D:
I absolutely love the design of Aiimages, and as a Unity dev I'm still planning to integrate it in to some of my projects. In the Unity space it's still IMHO the best SD tool available, by far, and I've tried them all (except Keijiro Takahashi's recent addition for the Mac folk). Props are due for the work you've put into it!
My waifu is fairly experienced with Python. If you have a sense of what would be involved to scan and rewrite the history file, would you be ok with sharing your pseudo-code thoughts on what steps would be required in Python to achieve the history restoration?
If we have any luck achieving it, we'd of course share the scripts with the community in case anyone else encounters the issue.
Many thanks again Sunija! <3 x
Hi! When I try to load the image I get this message:
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 16.00 MiB (GPU 0; 6.00 GiB total capacity; 4.34 GiB already allocated; 13.56 MiB free; 4.41 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
>> Could not generate image.
>> Usage stats:
>> 0 image(s) generated in 3.99s
>> Max VRAM used for this generation: 4.66G. Current VRAM utilization: 4.66G
>> Max VRAM used since script start: 4.66G
The image doesn't load and it stays grey :< Is there anything I can do? I have GeForce GTX 1660 Ti
Sounds like you're low on VRAM. :3 There are several things that you can try:
1) Close InvokeAI and check how much VRAM is already used up by the system in your task manager. Try to close stuff that you don't need and check if it lowers the VRAM usage.
2) Disable the NSFW checker, as it also needs some VRAM. You have to restart InvokeAI so the setting is applied.
3) Disable the in-progress preview, because it also needs VRAM. You can disable it in the tool on the cogwheel in the top right -> Display In-Progress Images -> None
4) There is a --free_gpu_mem option that will save VRAM. It will slow the generation a bit, though. To activate it, you have to open invokeai_2_3_0_standalone\invokeai\invokeai.init in a text editor and add the line after one of the upper rows. But I'd first check if the rest works. :3
This can be (random) cause when you set image size out of range. Try setting the image size between these number (512, 768, 1024). When image size out of these number sometime it work, sometime it won't. Make the prompt longer also fix this problem. And finally if it doesn't fix, your VRAM might already full.
Just a side note, after reading about Windows antivirus below. Bitdefender saw two of the Python files as Ransomeware. I added then them as OK and the install continued as normal.
HI. Unfortunately, when I launch Invoke AI I get an error, specifically PermissioneError WinError 10013 Attempt to access the socket in ways not allowed by the respective access permissions. I don't know what to do unfortunately
Can you try to run it as an administrator? :)
(Rightclick -> Run as administrator)
If that doesn't help, poke me on Discord (Sunija#6598) so we can look for a solution. :)
Thanks for availability, my discord is ElPolloLoco#0042, I tried to run it as administrator, disable windows defender and windows firewall but it still doesn't work
Hey! Great job, it's super convenient and easy!
Can we expect LoRAs support soon? If yes, then roughly when? Around April? Thank you for your work!
I keep getting a Down Load Issue at 954Mb, I've tried 3 times. It could be a local issue.
If you download via the itch.io/app then you should be able to resume the download. :)
You were 100% correct, downloaded and installed beautifully!!
I'm having a problem with the easynegative embedding and as the only thing I found about this is another comment, i'm trying my luck too:
>> Notice: embeddings/easynegative.pt was trained on a model with an incompatible token dimension: 768 vs 8
The embedding does not show in the list and does not seem to do anything when I try to use it anyway. (No ">> Preparing tokens for textual inversion")
Guys, if your trying to use an AMD gpu it will not work on windows (for now). But if you want to use stable diffusion i cant recommend Google colab (Stable diffusion) enough.