Failure to generate image, CUDA error | Invoke | Page 1

rocky vortex Dec 14, 2022, 12:43 AM

#

I've been generating images for a while on this same PC since the 2.2.3 release and then one day last week it stopped working. I spent a lot of time installing, uninstalling and reinstalling. I get this error whenever I'm trying to generate images now and I don't know what to do.

I've recently installed the 2.2.4 release at it appeared to work but in the WebUI it goes into a preparing loop and then in the console it gives me the attached error.

Also where do I enter the CUDA_LAUNCH_BLOCKING=1

unrelated/related maybe, did InvokeAI break my card? and that's why I'm getting the CUDA error?

#

also this happened using the anything v3 model but also happens with the stable-diffusion-1.5 model

drowsy radish Dec 14, 2022, 1:17 AM

#

What's a prompt you're using?

#

There seems to be an issue currently with longer prompts than can be actually used by the system. Looking into it now as a matter of fact 👀

rocky vortex Dec 14, 2022, 2:37 AM

#

It was a long prompt, it had a bunch of Danbooru tags in it as well as maybe a line or two of negative prompts using the [] and a variety of important prompts using () and (()). Sadly I’m not at that computer right now, oh but I did save the prompt to my cloud, here’s a screenshot

drowsy radish Dec 14, 2022, 2:39 AM

#

Yep - Ok! There's a bug right now on the unified canvas specifically with prompts that do not fit (and are truncated)

#

A few other notes -- () and (()) are auto conventions, and don't work in Invoke

#

Should use the ++ and -- syntax for invoke - We also have other advanced features (and blends, which are effectively how auto gets around the 77 token limit) which you can find the prompt syntax docs.

#

https://github.com/invoke-ai/InvokeAI/blob/development/docs/features/PROMPTS.md#prompt-syntax-features

GitHub

InvokeAI/PROMPTS.md at development · invoke-ai/InvokeAI

This version of Stable Diffusion features a slick WebGUI, an interactive command-line script that combines text2img and img2img functionality in a "dream bot" style interface, and...

rocky vortex Dec 14, 2022, 2:41 AM

#

Gotcha, I wasn’t using the unified canvas though just text to image. Really () doesn’t work, but the prompt guide for invoke said it does, at least it did with (())+++ when I first installed 2.2.3

drowsy radish Dec 14, 2022, 2:42 AM

#

think of () as just a container in our syntax

#

no weighting effect

#

() and ((())) are the same

#

()++ is weighted higher than ((()))+

rocky vortex Dec 14, 2022, 2:43 AM

#

Hmm, so I need to use the + and - with the containers now?

drowsy radish Dec 14, 2022, 2:43 AM

#

always did.

#

well

#

you dont HAVE to

#

but if you dont it applies to only one word

#

so
(like this)+ works (like)+ (this)+

and like this+ works like (this)+

rocky vortex Dec 14, 2022, 2:46 AM

#

Okay, thanks, I’ll need to give that a try, so (()) and ((())) don’t work but I can still use () as long as it has a + or - attached?

drowsy radish Dec 14, 2022, 3:01 AM

#

They "work" they just arent necessary 🙂

#

you can nest terms (this (and this)+)+ work like (this)+ (and this)++

rocky vortex Dec 14, 2022, 2:21 PM

#

okay, the following are new attempts and I tried it three times, modifying the prompt so the package had a + at the end (3 screen shots), modifying the prompt to be much shorter and only including 1 negative prompt in [] (1 screen shot) and finally, using no negative prompts and a shorter prompt. All of them done using the SD1.5 (I'll be uploading the SS's in order:

#

#

each time I had to reload the browser window to start another test

drowsy radish Dec 14, 2022, 3:01 PM

#

Your prompt is too long and hitting a current bug in the system

#

It should just ignore it when your prompt is too long but atm it’s choking. Try shortening it up a bit

rocky vortex Dec 14, 2022, 3:06 PM

#

I did in the second and third try, still broke, but when I reloaded the invoke.bat and started again it worked for a while. I kept upping the complexity, adding some extra words, increasing the negative prompts and it eventually broke. It's kinda annoying. And the Anything V3 model takes ages to load again.

#

not that the model taking forever is invokes problem. (wish they loaded as fast as a1111)

#

Strange thing about the model load times. I had it on a hybrid drive that I was pulling the models from and it loaded much faster in 2.2.3 than on an NVME SSD with 2.2.4

#

Anyway, I’ll use shorter prompts until they can fix whatever issue is causing the breakages. Hard to do when using danbooru tags in addition.

drowsy radish Dec 14, 2022, 3:26 PM

#

rocky vortex Anyway, I’ll use shorter prompts until they can fix whatever issue is causing th...

Another method that can work to add complexity is to use prompt blending

#

("Prompt 1", "Prompt 2").blend(1,1)

#

this is effectively how Auto does > 77 tokens, but does it behind the scenes (so you lose control on what parts are blending)

rocky vortex Dec 14, 2022, 3:30 PM

#

When you say 77 tokens. Does that mean 77 words or 77 phrases or something else?

drowsy radish Dec 14, 2022, 3:33 PM

#

this gets into the bits of SD that are a bit less intuitive

#

https://beta.openai.com/tokenizer this is not a 1:1 with SD but a good conceptual example

OpenAI API

An API for accessing new AI models developed by OpenAI

#Failure to generate image, CUDA error