Controlnet QR Code Monster For SD-1.5

@achiru I get that. Just figured it was worth pointing out, since my default answer to people who say "I don't have a QR scanner" it to tell them that their phone's built in camera app can do it.

@achiru I haven't seen a QR code generator that lets the user choose 16px for the module size. The closest I found is https://keremerkan.net/qr-code-and-2d-code-generator/ which has 15.

So this shouldn't be used with txt2img, but img2img with a blank gray image as the starting point? It doesn't seem possible to start with a solid gray image in txt2img.

@achiru So like this? https://i.imgur.com/M3gRXE0.png

@achiru Thanks for the link to the generator that allows customizing module size. The QR codes it makes aren't transparent though, so the gray background would be completely covered right?

Hi, great work! I would also be interested in understanding the bit with the "gray" image in the background. Is it pasting the qr code on an image with gray background ? Why exactly is that ? Thank you.

@achiru Hmm, I'm unable to get this to work... In the ControlNet panel, what Control Type should I choose? All, Canny, Depth, etc

@achiru Thanks a bunch!!

@achiru I tried to get the image metadata and I am not seeing it. I am more used to getting the settings from .png, but I looked at the exif info on the .jpg and didn't see anything.

I am currently using 1.2 for the controlnet weight and .3 for start an .95 for end.
I am using realisticVision_V40 with the included VAE.

I have gone as high as 1.6 for the weight and as low as .15 on the start.

I am getting better images now. I guess it depends on the prompt. Some don't see to incorporate the QR code as well.

I have been using a 700x700 qrcode for a 768x768 image, I looked at your qrcode and thought that was what you did. Also I started using the QRcode generator that you recommended.

@achiru I'm using Beta options, so the documentation I can find is here: https://github.com/easydiffusion/easydiffusion/wiki/ControlNet

You'll need the control_v1p_sd15_qrcode_monster.yaml file from here https://huggingface.co/monster-labs/control_v1p_sd15_qrcode_monster/tree/main place it next to the controlnet file and name it the same (both need the same name not the same extension)( example: qr.safetensors & qr.yaml will work just fine)
Then just set to Euler A, 50 steps, 7 guidance, no negative prompts (at least for my experience) and it should work just fine. Here's one I made: https://ibb.co/nCSWJNT hit me up on Discord for help

@_o_w_l thank you

That's because autocast isn't supported on Apple's backend, amazingly enough.
As for fixing it, you usually need to change the attention model you're using when you start comfy... usually.
I think Quadratic (the default) was the only one that worked for me when I had this happen on the DirectML... until I added a hypernetwork too and that one called autocast from another attention function and I finally deleted all the autocast blocks.

@GnomeExplorer I don't think I would know how to do that myself but happy to watch any tutorial if you have one, thanks

@kameleo333 It should be fixed now anyway but if it crops up again, some command line options you can try when you run main.py since autocast is normally used to upcast fp16 -> fp32 during attention and CLIP :
You can try the 3 different attention models one at a time and see if one of them gets through. The difficult part is that some custom nodes added their own attention types that won't respect the datatype flags or don't upcast but you should be fine with just the controlnet.

--use-split-cross-attention

--use-quad-cross-attention
--use-pytorch-cross-attention

If that doesn't work, try
--fp32-vae --force-fp32 (mac should run in fp32 normally anyway the way comfy configures it... if you don't specify , but it might be using fp16 for something)
the above should have done it but who knows, again

or
--dont-upcast-attention prevents the attention models from using autocast (I think) which would eliminate this the next time a new attention type shows up. I've seen quite a few in the past 3 months, it's evolving fast.

If those don't work file a bug on the git :-)

I hope so as well

@achiru Awesome work mate! Can't wait for SDXL version! You are a legend!

Here is a tutorial,
Just use the latest version.
Now it's V2 =)

https://www.youtube.com/watch?v=nFfHEmNu0Zk&t=553s

I had some success implementing it as a Controlnet node reading model from the checkpoint and throwing the control_net to an Apply Controlnet (adv) node that is also receiving the image of the qr code from a Load Image node. From that, you combine the Apply Controlnet node with an empty latent image into a K Sample and voilá. You will need two clip text encodes between your load checkpoint and the Apply Controlnet node, of course, for prompts. I hope it works for you as well

You have to go for simple styles. Anything too much and most QR scan apps won't be able to take it. 99% of the created images you see here are unusable for scanning but they do look pretty

Thanks - what does this file do?

i need this too

Use quik QR art

It's a shame there's no standard for hashing of the password in WiFi QRs (or at least I couldn't find one when I was generating a bunch last year).

The codes for wifi will usually end up fairly long because of the length of text and error correction so to reliably get codes working that didn't overwhelm the generated image my strategy was to set up one controlnet that at a strength < 1 and a range of about 75% of the steps for the original image... the output from this can sometimes scan but most of the time if you've got a complicated prompt or a bunch of LoRAs some of the important areas will change into objects too light to provide constrast.

Then upscale it with a NN latent upscaler to avoid a VAE messing with contrast and run a second QR monster controlnet on it for about 50% of the second ksampler @ strength of 1.05-1.15 with a noise reduction level of something around ~33-55% depending on the CFG and lora chain. Running it for 50% reestablishes any dark areas that got obliterated by the character lora and prompt during the low res generation just enough that the remaining steps will keep them around but fade them a bit or transform them into other objects and you'll get something that scans about 90% of the time but isn't immediately obvious as a QR.

It's a little slower this way from running two controlnets but otherwise you end up with too many generations that need to be manually fixed after upscale and the QR tends to get really obvious if you resort to that.

https://www.reddit.com/r/StableDiffusion/comments/18aoqs0/complete_qr_monster_v2_workflow/

@zryn the link is linking to SD1.5. I came here to find out if there's a Flux or Illustrious qr monster workflow.