fyi that's you issue not a model issue.
Do what the instruction says.

I mean prompt examples?..

Images here are 'cleaned' when they are stored by Civitai and they don't have metadata anymore. In other words, they cannot be used to recreate a ComfyUI workflow.

On top of that, Civitai works very well with A1111 but not well at all with ComfyUI or other UI and has difficulty to read metadata once you deviate from the simplest workflow.

Watch Scott Detweiler's series on YouTube. He is the best teacher so far when it comes to teaching Comfy from scratch. Watch the first few videos of his comfy series and you'll know how to create your own custom workflow.

After that, you'll never go back to Automatic1111.

@ImAbbieKitten This. This is where I learned and still learn to use ComfyUI. A great advise for a beginner..

@KandooAI I'm still keen to play with it! Would you say it's better than the diffusers/stable-diffusion-xl-1.0-inpainting-0.1? I'm loving Juggernaut XL v7 for inpainting, it's just that it has lots of coherency issues (as non inpaint models do 😊)

Also, how long can i expect to wait, if you know?

@KandooAI Yes please! I would love to start using it - even if it's slightly better than the Stability one, there are no other inpainting models out yet AFAIK so it would be a massive help to lots of people 😁

Regarding the resolution is a SDXL model so it supports the SDXL usual resolutions like, for example, 1024 x 1024, 1152 x 896, 1216 x 832, 1344 x 768, 1536 x 640

Merge it yourself in Comfy UI, it works quite well, 8 steps for decent output at 1024x1024

Description says that version 7 VAE is baked in, so try not using VAE.

I'm also having this issue with the staircase/checkerboard. Using Auto1111 and the standard SDXL VAE seems okay, although I haven't done much testing.

@davz yes initially I was using the v7 default vae which is what gave me those strange artefacts, but this didn't happen in v6. So I'm thinking maybe the author baked in the wrong vae by accident during the making of this model.

@majorromp975 yes same issue on comfyui. To fix the issue, I had to use the xl0.9 vae or just borrow the vae from v6 by having two checkpoint loaders in my workflow.

@jason_chan_12790 I use ComfyUI and I had never any issue with any version.

I don't load any specific VAE, I just link the model to the VAE points and I am done.

Experiencing the same issue with the v7 model (but not v6) in both ComfyUI and InvokeAI. Using a different VAE fixes the issue for me too.

I'm fairly new at this so I will likely get this wrong. But I had a similar problem (A1111) and it turns out I had manually set the VAE in settings to another VAE. Putting this to automatic fixed everything.

@outlier_z It looks like they are doing something very similar but in Comfyui.

They have probably a 'Load VAE' node somewhere and they insist in using it when everyone including the Model desciption tell them to not do so, because the VAE is Baked in it.

Just connect the 'VAE Decode' directly to the node 'Load Checkpoint' throught the VAE point and that is.

There is no need to use a 'Load VAE' node and it can be dropped. You can't have it easier.

To clarify, I am experiencing this issue in ComfyUI when connecting only the Load Checkpoint VAE point directly to the VAE Decode VAE point. No other Load Checkpoint and no Load VAE nodes are present in my workflow so definitely no other VAE being used by accident.

@kyna Do you have any message in the command prompt window? Just before the process starts and during the run?

@Aerth Nothing that stands out to me as an error; the console output is identical between runs with v7, which shows the image artefacts, and v6, which doesn't:

got prompt

model_type EPS

adm 2816

Using xformers attention in VAE

Working with z of shape (1, 4, 32, 32) = 4096 dimensions.

Using xformers attention in VAE

missing {'cond_stage_model.clip_l.text_projection', 'cond_stage_model.clip_l.logit_scale'}

left over keys: dict_keys(['cond_stage_model.clip_l.transformer.text_model.embeddings.position_ids'])

Requested to load SDXLClipModel

Loading 1 new model

Requested to load SDXL

Loading 1 new model

100%|█████████| 50/50 [00:08<00:00, 5.98it/s]

Requested to load AutoencoderKL

Loading 1 new model

Prompt executed in 11.53 seconds

gc collect

Jaggies on contrasty and saturated contours, too. Please don't bake in the VAE or use a non watermark one or SDXL without watermark?

Here is what they did with sdxl 1.0
https://hforsten.com/identifying-stable-diffusion-xl-10-images-from-vae-artifacts.html

Please use the VAE from 0.9 or the fp16 Version above for v7 please, its a great version, but the watermark is a bugger. red channel seems affected most.

To add to my previous comment, this issue persists across A1111 and Comfy. I don’t see this happening with the previous version or other checkpoints/models with my current setup.

Edit: this is when used with Img2Img

And just to be clear Juggernaut’s amazing, my goto checkpoint for most things, thankyou for making this available KandooAI

What watermark are you talking about? I've seen no such thing, using this model.

Could you please give us some examples of such constrasty and saturated contours looking rather blocky?

https://hforsten.com/identifying-stable-diffusion-xl-10-images-from-vae-artifacts.html

@alecc What the article says (very interesting by the way, thank you) is the SDXL model doesn't have those artifacts but they are incorporated during the Decoding and "However, because the Stable Diffusion is open source and watermark is not a part of the neural net model and is applied afterwards to generated images it can be easily removed from the program. In fact, two of the most popular UI's A1111 and ComfyUI don't include this watermark and it is not present in the majority of the Stable diffusion images found in the wild."

Did I miss the point?
I don't say that aggressively but by curiosity as I am really interested.

@Aerth Yes, the blocking happens in the VAE, so by using a different VAE in Comfy and A1111 we can avoid it, right? But the VAE embedded in Juggernaut has it baked in, so in Invoke for example, where I can't change the VAE if its already baked in, I get those jaggied watermarks in v7. Other comments here are referring to the same problem.

OK, I read in another comment v6 didn't have the issue, so revised that. But could you maybe bake in the 0.9 VAE or use the fp16 model mentioned above to avoid the blocked jaggies? Going nodes in invoke to swap the baked VAE just raises complexity.

Maybe just don't bake in the VAE, then we can choose?

Thank you for posting this link, very interesting read! I've been experiencing those exact artefacts with the V7 model in ComfyUI and InvokeAI and have been tearing my hair out trying to figure out what was causing it! Very annoying having to pipe in a different VAE every time I use this model (which is a lot, its a beautiful model!) just to get clean outputs!

@kyna Yes, I hope @KandooAI sees the watermark and what we mean when comparing the images produced with VAEs from 0.9 vs 1.0 or with SDXL vs the fp16 version with the same output I linked above. I really cant work with this models blocky VAE, the model is great though of course.

@KandooAI Wonderful! So the 0.9 VAE is not the way to go? It has a different license, hasn't it? Otherwise the same? Why bake it in at all? We could choose ourselves, no?

@KandooAI Looks perfect, thanks!

Good question. Thank you.

Did you also copy the seeds?

@gfour Yup everything...better! I reinstalled whole automatic1111 coz of this, and it's still give me different image, similar, but not the same.

Like, how different? It's next to impossible to get identical results. If you're using XFormers the results are apparently hardware-dependent. There are other differences as well.

Also make sure your image resolution matches exactly. Different resolutions (and in particular different aspect ratios) can result in wildly different images. It's very unlikely that the images uploaded here are the original size. They're upscaled. That means to make them identical you need to know the original resolution, upscaler used, upscaling steps, upscaling factor, and the denoising strengths. Those settings aren't given in the post.

The only thing I can guess is that this batch seems to be generated at 768x1152 and upscaled by 1.5 (the final images are all 1152x1752, and 768 is a normal starting width). Other than that there's no way to tell what settings were used.

@certifiedDoc 

Of course that i can get same image with this settings, but something with mine automatic1111 is not right...

beautiful lady, (freckles), big smile, blue eyes, buzzcut hair, dark makeup, hyperdetailed photography, soft light, head and shoulders portrait, cover

Negative prompt: (worst quality, low quality, normal quality, lowres, low details, oversaturated, undersaturated, overexposed, underexposed, grayscale, bw, bad photo, bad photography, bad art:1.4), (watermark, signature, text font, username, error, logo, words, letters, digits, autograph, trademark, name:1.2), (blur, blurry, grainy), morbid, ugly, asymmetrical, mutated malformed, mutilated, poorly lit, bad shadow, draft, cropped, out of frame, cut off, censored, jpeg artifacts, out of focus, glitch, duplicate, (airbrushed, cartoon, anime, semi-realistic, cgi, render, blender, digital art, manga, amateur:1.3), (3D ,3D Game, 3D Game Scene, 3D Character:1.1), (bad hands, bad anatomy, bad body, bad face, bad teeth, bad arms, bad legs, deformities:1.3)

Steps: 35, Size: 768x1168, Seed: 3218107096, Model: JuggernautXL_RunDiffusion_V7, Version: v1.6.0-2-g4afaaf8a, Sampler: DPM++ 2M Karras, CFG scale: 7, Model hash: 0724518c6b, Hires steps: 15, Hires upscale: 1.5, Hires upscaler: 4x_NMKD-Siax_200k, Denoising strength: 0.3

@certifiedDoc you just need to click that little "i" icon under image and click "copy generation data" and you'll have all image settings, try it.

@KandooAI I'm using Colab version of automatic1111 named thelastben. Two days ago it gives me results like yours and I didn't change anything in settings maybe author of Colab change something... anyway thx for ideas what could be wrong. I'll give feedback when I solve the problem.

@KandooAI it doesn't work, i'm out of ideas what could be wrong -_-
I installed xformers on colab like that:
!git clone --depth 1 https://github.com/facebookresearch/xformers.git %cd xformers !git submodule update --init --recursive !pip install -r requirements.txt !pip install -e

ENSD value is "0" in settings.

This is mine webui:
version: v1.6.1  •  python: 3.10.12  •  torch: 2.1.0+cu118  •  xformers: 0.0.23+3f74d96.d20231024  •  gradio: 3.41.2  •  checkpoint: 0724518c6b

20 minutes? What are you trying to do?

@KandooAI Maybe my settings are wrong, but I copied the same settings from the redhead image that you posted and ran it, but it was my first time running XL so I'm sure I did something wrong.

@sylentwulf81 Perhaps my comment will be useful to you, I also have a 3070Ti and the same problem in Automatic1111, I tried to install another gui (Fooocus 2.1) and the problem is solved, pictures of the same format are created in 1-2 minutes

@zaqhgh I'll try that, I know a lot of people on here use A1111, but based on your comment I watched a video on Fooocus, and it looks amazing. I'll give it a try and post an update.

@sylentwulf81 Also try out adding parameters to your A1111 webui based on this:
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimum-SDXL-Usage

Basically, if you have 8GB VRAM Nvidia card, you should open your /webui/webui-user.bat, and edit the COMMANDLINE_ARGS line to say this:

set COMMANDLINE_ARGS=--medvram-sdxl --xformers

This helped me get from ~20s/it to around ~4s/it depending on resolution. It also uses less VRAM so it fails less.

@Stryp This helped a lot! Generations still take a lot longer in A1111 than in Fooocus, but it cut the render times down to about 1.5m - 2m. Thanks for the tip!

yes for me to it takes about 30 min for a single image. i tried it using fooocus and it seems to be a little fast. usually it takes me about min to generate. know i am using aftermath. but thanks again kandoo ai