1st :)

Love your work, thanks the release!!

This is just magical

Here's the famed new fancy model everybody is talking about, great work!

when i swichting to this ckpt,

RuntimeError: expected scalar type Float but found Half

idk how to fixed it

I am getting this error when trying to generate

<urlopen error [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed: self signed certificate in certificate chain (_ssl.c:997)>

**It seems it is related to the LDSR Upscaler.. I can use ESGRAN-4x fine but the output isn't optimal, and if I turn it LDSR down to 1 I can use it, but it also gives me a deformed generation.

Since past 1 month, I've been only using Illuminati Diffusion. Goes with everything.

no need for a config file?

Very high quality and flexible. Good job.

The major quirk I've noticed after some testing is that human faces look a bit wonky and unappealing.

modules.devices.NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check. （how to fix？）

why doesn't LORA work in your model?

Liked, so i can downlaod it when i'm done gaming and toy with it XD

Excellent model. Can any one tell me if he was able to use Controlnet with this model?

Is there a way to get this to not do gaunt looking faces? It tends to lean towards supermodel sort of styles: IE. super thin, gaunt, over exaggerated, angular faces, pointed chin, no chins.

would be better if this model didn't have tendencies to create product shot-like images

Why does it need 3 different negative embeddings? Care to explain what they do? What happens if we don't use them? Can this model also generate "normal" images or only high contrast?

The discord link is expired!

Clearly trained with midjourney images, it outputs very stupid faces, plastic, elongated, anorexic, just like MJ does

Is this model was training with Midjourney Images?

Some of my results with the 3 Negative Embeddings and the model, the results look so similar to Midjourney, I'm a experience Midjourney user & Stable diffusion, so it's really cool the mimic of style, Or it's just this contrast training that give that feeling? What do you think people?

Can we have a 1.5 version?

A model with a very particular style that I'd hesitate to tag as "general purpose". A detailed prompt will result in the same human face across every seed, which might be good if you are looking for consistency, but not if you want variety.

Are VAE's a thing with 2.1?

I only got black images :(, anyone know this issue?

not funtion!

I get this error when trying the sample knight bloodborne image:

mat1 and mat2 shapes cannot be multiplied (154x1024 and 768x96)

any idea how to make it work with controlnet i keep getting following RuntimeError: mat1 and mat2 shapes cannot be multiplied. It doesn't matter what resolution or option i change it doesn't work together

Any idea how to leverage embeds in colab?

Very pretty results....But this model has a serious eating disorder. All women looks facial features that look very skinny, and muscly.

When I try to use my Lora I get a runtime error.

Would be VERY helpful if you provide context for the fixer embeddings you provide - and clear up the runtime issues some users see with Loras. We can't keep up with everything (intellectually) - in the SD space, c'mon =)

Awesome 🤩

You will add the model to HuggingFace?

what VAE is better to use with that model?

now we need someone to make it re-learn anatomy by training it on nudes in order to completely fix what stability broke

Hi, how to install the illuminatiDiffusionV1_v11.yaml file?

Looks fantastic! Any plan to have a diffusers version? Thanks a lot to share this with the community.

It seems like this model is biased toward dark/black background. I tried to add white background to positive prompt, and dark, black to negative, but it still generates dark background. Anyone knows how can I improve this?

How would it be possible to train a Lora based off of this? or would that be out of question for specific subjects?

Which values of offset noise did u use for training?

Default 1.0

Just about everything in this model is aesthetically pleasing. I'm wondering did you also use some sort of image classifier to drop out bad aesthetic images?

How do I load them negative prompt files?

I get this error:

NansException: A tensor with all NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.

I have the yaml file in the same folder as well as the 3 embeddings installed.

Any ideas?

Somehow or... maybe I'm missing something here,
whenever I switch to your checkpoint, all my other embedding files disappear(also become unusable too) but only your "nartfixer" show up?

Why can't I merge this model with other models?

hires fix not working.. how to fix this problem?

idk why this is popping up but i installed all files and still get this error: changing setting sd_model_checkpoint to Illuminati Diffusion v11.ckpt [cae1bee30e]: AttributeError

Traceback (most recent call last):

File "E:\StableDiffusin\stable-diffusion-webui\modules\shared.py", line 568, in set

self.data_labels[key].onchange()

File "E:\StableDiffusin\stable-diffusion-webui\modules\call_queue.py", line 15, in f

res = func(*args, **kwargs)

File "E:\StableDiffusin\stable-diffusion-webui\webui.py", line 146, in <lambda>

shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights()))

File "E:\StableDiffusin\stable-diffusion-webui\modules\sd_models.py", line 488, in reload_model_weights

state_dict = get_checkpoint_state_dict(checkpoint_info, timer)

File "E:\StableDiffusin\stable-diffusion-webui\modules\sd_models.py", line 262, in get_checkpoint_state_dict

res = read_state_dict(checkpoint_info.filename)

File "E:\StableDiffusin\stable-diffusion-webui\modules\sd_models.py", line 248, in read_state_dict

sd = get_state_dict_from_checkpoint(pl_sd)

File "E:\StableDiffusin\stable-diffusion-webui\modules\sd_models.py", line 197, in get_state_dict_from_checkpoint

pl_sd = pl_sd.pop("state_dict", pl_sd)

AttributeError: 'NoneType' object has no attribute 'pop'

What is this error?

RuntimeError: mat1 and mat2 shapes cannot be multiplied (308x1024 and 768x320)

It's a model that can make some interesting outputs, but I think there are a few issues with it that need to be addressed by the model maker.

Firstly, this model has a propensity to create yellow/blue green lighting (especially yellow) and leans heavily toward a very similar and almost singular lighting/color scheme. It's almost like whatever images it was trained with are heavily bleeding into every output. Yes you can prompt around this, but it really isn't desirable as that does limit its ability to make it suit to an individuals style, and trying to prompt out the yellow can also cause some wonky colors, like things that should be yellow, looking more like a yellowish green, and things that shouldn't be yellow, having yellow for no reason.

Basically, if someone is using this model, it's easy to tell, because outputs look a lot like if you had a photographer with the same style, which means it's going to go out of style real quick as people find their outputs of a variety of subjects just look too similar in style, even if they're entirely different subject matters. Again, you can counter this somewhat with additional prompting, but that will also affect the output by having to add words you otherwise wouldn't need, so it's a bit of a crutch.

In many cases the way it does lighting can be too dark or too high contrast. I understand that darkness is the point, and again it can be prompted around somewhat, but I think it needs some kind of tweaking in this regard.

It also has a propensity to create something almost like compression artifacts in dark areas of images, like a JPG saved on a low setting. I don't know what causes this, and it's not always present, but I think if they used JPEG's for training, the dark images may have not contained enough data and had compression artifacts from JPG not being able to display the full range of tonal gradation, which seems to have gotten baked in with the training.

My suggestion is training should be redone with at least native 16 bit, if not 32 bit images, especially because it's a dark focused model. One has to keep in mind during training that the AI see's the raw data of an image. So even if it's too dark to see artifacts normally with the naked eye, the training will see it and bake those compression patterns into how it functions.

I know this is built on SD 2.1, which probably didn't have high bit images in mind when training, but whatever additional images might have been used to create this model should be scrapped and similar but minimum native 16 bit images used in their place, which should vastly increase the output quality of the model in general.

The Automatic 1111 web UI folder was not found，The embeddings folder in the root directory can not be read

Very good model, but there is a common tendency of fixed small eyes and high cheekbones on Asian/japan/korea/china faces, hope the next version can improve.

You ever going to make a 1.5 version?

The config file doesn't work, I'm using INVOKE and I tried everything..... I have the embeds taken, I put the config along the model file i put it also in the specified folder of invoke and nothing works. the model doesn't load...

ERROR: LatentDiffusion.__init__() missing 1 required positional argument: 'personalization_config'

CONTROLNET doesn't working with this mod，I download the sd2.1 controlnet,No errors report, but it doesn't work

I am also unable to get the model to load in Invoke, following all the normal model loading procedures required from the UI without success ??

Any ideas would be appreciated.

So we cannot use LoRA such as KoreanDollLikeness? I got the error RuntimeError: mat1 and mat2 shapes cannot be multiplied (77x1024 and 768x128)

why i can not use this model.When I render with this model, the rendered image is all black and nothing, right

Pretty results, but the model has a tendency to make very long people, even on 768x768 square sizes. It's also very dark, not very illuminated in fact.

For those having issue, here are the trigger words if its not working by <filename> for you:

nfixer.pt - nfixer-271

nartfixer.pt - wer

nrealfixer.pt - nfixernext

Does anyone have instructions as to wether these are supposed to be use one at a time or all in the negative prompt at the same time? What is supposed to be the different between them?

So does this mean it doesnt work with SD 1.5?

I love this model. I'm wondering if you could add some explanation on what each of the negative embeddings affect. I have played around with adding all, or just one or two. Like, if I put just nartfixer, I would assume it has something to do with non-photographic images. Thanks for all the work that went into this!

bro for some reason i cant replicate any pic i get really bad results even with just putting the image in png info and sending it to txt2img

Congrats Civitai!! Republish all the images of this model is a great decision!! Thanks!!

Why did this model pop up once again the new models list?

when new update? any plans for SDXL?

Can you please enable the model to allow it to be used for civitai's online image generation? Thank you for your consideration.

Greetings, can you elaborate on the noise offset value you utilized during training? I know you are considered a specialist!