Btw, in case you wanna add it to your readme, the "Sampler: DPM++ 2M Karras" (which is an A1111 label) should be set up as follows in ComfyUI:

- In KSampler's settings, configure it as follows:

- sampler_name: dpmpp_2m

- scheduler: karras

ComfyUI lets me switch VAE.

Hope it's not some track code to mark content

cant type nothing there automatically closing after i pressed a key

If you use Automatic1111 - first update it to latest 1.6 version, then go to your stable diffusion folder, and find the \venv folder. Rename it to something like "venv_backup". After this, run A1111 again, and pay close attention to the cmd window. It should automatically create a fresh \venv folder, and install all the crap it needs for SDXL to run. Worked for me like a charm, and I had the exact same error code as yours, among a few others.

@3drazer1911344 still wont work imma go crazy all i wanted is just testing this model i ve tried everything it still says no module xformer founded proccesing without it :((((((((((((((((((

Open a terminal and run the following command to install the specified version of XFormers:

pip install xformers==0.0.16

@jyankienn742  don't open it, right click and open with notepad or edit.

@KandooAI I have an other questions how they make videos because I don't find in my profile just create image

there should be a little folder under the pictures click that... a quick note I shortcut the folder that way I don't make copies & it's tidier that way.

@KandooAI Thanks for the detailed response. You're right about 5 steps being the perfect number, which technically goes against the 1 step feature. Also, the CFG is very limited (4 is the max, and most of the time it's burnt), so your prompts need to be solid or you'll be fighting for details. Yes Turbo has limitations and it needs work, but my god the latent upscaling looks amazing. Check out the images in my ComfyUI workflow, all latent upscaling.

https://civitai.com/models/227882

@KandooAI You're such a strict man. RESPECT~but I'm still eagerly awaiting to see the release of the turbo version.

@KandooAI Dreamwalker! can't find it!

Try adding --medvram to your Commandline Args. On an 8 year old computer with a GTX1080 card, I can generate 768x1024 images in less than a minute. Without that commandline arg, it takes upwards of 10 minutes.

You can also try Comfy UI.

I'm on a GTX 1070 and I can generate an 768 x 1152 image in about 60 seconds (DPM++ 2M Karras), with 20 steps ( although the first time I use the model, it takes about more 20 seconds, to load ( with a SATA SSD).

And I don't seem to have any VAE errors with the original version 7 ( before the VAE fix version ).

all OK ;)

i changed the option of the model, (i have never used that . . .)

@OmniNikon try using adetailer to get better hands. Even just putting "hand" in the adetailer prompt with the yolo hand mask can generate good results. I usually do two or three passes with adetailer with decreasing denoise levels.

@KandooAI Wow! Thanks for the reply. Glad to hear that its a priority. Can't wait to try out v8 when it's ready~

make it yourself

@CoolToast this is not Reddit so fuck off

If the output looks identical to you, that is great. The VAE fix version is an fp16 version of the default fp32 sdxl vae. This means that the network weights are scaled down from fp32 to fp16 which takes less vram and should be faster at almost no quality loss which you have just proved by not being able to see a difference. I do see a slight difference sometimes.

Agreed.. I've tried so hard not to get blurred backgrounds, but I can't find a way.

See https://civitai.com/models/133005?modelVersionId=252902&dialog=commentThread&commentId=315334&highlight=316258

Thanks

getting the same thing - looks good up until it finishes and then gets really high contrast/bright colours. switching to other model is fine

What sampling method and scheduler are you using? Try dpmpp_2m Karras. And make sure your denoise is inline with the model.

I've had this problem occur when accidentally using an improper VAE.

Same issue here...

Just disable (Automatic or None) VAE settings in Settings tab :)

I have this too!

Just disable (Automatic or None) VAE settings in Settings tab :)

It is as it says, you need to install xformers. Either install with pip as it says, or edit your webui-user.bat file to include the following commandline argument:

set COMMANDLINE_ARGS= --xformers

Next time you run the batch file, it should install xformers for you.

@KandooAI Sounds good!

Let's check it out. Does Auto1111 finally support inpainting?! Email [email protected] so we can chat about it.

oh its not finally

you can see here

https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/13195#discussioncomment-7932146

https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14390

and the 4 files

you should update to 1.7.0

https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14390/files

icant tell any more cause im not a developer ;)

Note that inpainting with the regular JuggernauXL model works perfectly under Fooocus without any tricks because of the way they implemented it.

@joeuser12 

yes most normal SDXL models work with inpaint but not that good like old SD15 ... with this first try i get much better results ;)

... ill wait until KandooAI read this and maybe you get it

@RunDiffusion How did you guys make inpaint models before this? With comfyui, you get a model file that is the same as the base model, but it does have the inpaint feature. But it's not even the same size as the official inpaint model.

@sevenof9247 I didn't really understand your response. Neither Comfy nor a1111 1.7.0 come anywhere close to the inpainting/outpainting quality you get from Fooocus using this model. lllyasviel, who is the author of ControlNet, created a special inpainting controlnet model that is only available in Fooocus and it turns JuggernauXL (and any other SDXL model) into an inpainting/outpainting magic model that works way better than anything you can do with SD-1.5 inpainting models.

@joeuser12 

I didn't know that yet, thanks for the tip, I'll give Fooocus a try ... It seems to have even more functions and is a bit easier to use

Despite all this, it is unfortunately an extra program, again many GB load the program updates etc. ...

and what I have described is relatively new, the new original inpaint_sdxl model is only a week old ;)

and I am now trying to create new models e.g. with Jugger and others and the possibilities with the extensions in automatic1111 , inpaint-anything and segment anything ...

and controlnet in itself is a powerful toolbox, but I think only a view hundred can work with it. ^^

So how do you make the inpaint model. i use merge checkpoint for add difference and get a model though SHA256 that looks different. But it is exactly the same as the image generated by the ORIGINAL inpaint_sdxl model.

On the contrary, the inpaint model made with comfyui, which has inpainting capability though it has the same specs as the base model, generates a different image.

@suede2031691 

so if i remember, sry had some tryes... it was Enfugue, somwere was writen thats at the moment the only choice, after that i converted it with automatic1111 to fp16 and merged the VAE...

now it works with the patch much bertter than a normal SDXL model

@sevenof9247 Then I should be operating something incorrectly, auto1111 and enfugue get the inpaint model, they have the same sha256. I've only tested it at comfyui, and this model and the Official inpaint model will produce exactly the same image.

@suede2031691 

shame that civitai has no internal chat ^^

which two models do you mean ?

for me the sha256 are different, jugger and the offcial ...

and i use the model in a1111 in img2img -> inpaint with a mask and offical and jugger has different images

Sure u can

@denrakeiw That's how I've been using it. It s nice~

I'm just getting into training Lora, my first batch are pretty bad so I'm not uploading them to civit, but from my research, its possible to further finetune, or train lora from a finetuned model, but the best result would come if the unpruned (13gb version) of the model in question is used.

It turns out that V8 has been released over here, and there's no word on the inpaint model over there.

The author, Kandoo, has experimented with pornographic nsfw content before, but received feedback that it negatively impacted a lot of other content. So unless another branch is listed, it's just a lot of lora to get nsfw content.

But I've used multiple versions of Juggernaut XL, and they all work well with all types of nsfw lora. I'm waiting for a couple of models to be updated, and then I might be able to get an ideal nsfw model by merging them. It would require a lot of testing work though, or you could try it yourself if you have the time. The downside is the lack of some nsfw content other than porn, such as horror themes, and the lack of facial expressions on the characters, all of which puts me further away from the perfect nsfw image.

@KandooAI Is it turbo? Please say yes! :D

i was also checking constantly because i knew it was on the way, and then boom, here it is. I know it will be great.

@KandooAI I hope it will have huge amount of new clothing styles, NSFW poses and stuff :D

I can't either and that sucks, quess I'll stick with version 7.

This model needs more NSFW F YOU

@ArconSeptim lol you seem to be really touchy about it

Try applying some specific loras if you're after a specific fashion type.

Try different names, ethnicities, appearance features, professions, photography style... if you give a general description, you will get a general look.

Describe their features (eyes, nose, hair, face shape, etc). Your negative prompt is as important as your positive prompt. You can play words and phrases in there as well.

yep that's exactly correct! Just set the model to use VAE "none"