@KandooAI I'm perfectly happy using 7.5. Just thought I'd mention it. Keep up the good work!

@KandooAI can't wait for NSFW update, dude I will scan whole Playboy magazine from 1990s to now if needed haha

@KandooAI of course!

@wyxzddsjj919 it's open source model done by individuals alone, compared to massive corporations and their dedicated team, and we are still going in the right direction with models like JuggernautXL V8 with all these improvements

@KandooAI Do u have to use a refiner or something with it?? i can't seem to get better results specifically with this model.

Personal favorite
1216 x 832
832 x 1216

@KandooAI thanks you are kind

@KandooAI A bit better, thanks

Nsfw best u will lose 40% of customers

Hey check the VAE setting. Make sure it's not trying to inject a different one during the VAE Decode process.

try change output to 1024 x 1024 and disable upscaling, along with VAE suggestion.

github ->cumfyui impact pack

generate base image with ksampler, then feed the outgoing latent into this set of nodes:

load upscale model node -> latent upscale on pixel space -> ksampler -> vae decode -> save image. download whatever upscale model you want to use.

Try drawing realistic hands on your own and you'll understand, hands are and always have been very difficult, both for AI and for experienced artists.

@sken6825368110 I graduated from Pratt and yes those were my first lessons to learn in human anatomy :) And no it's not hard to draw them if u figure out the basics and lots of motivation. Dunno what your definition of experienced is though as none of my peers ever complained about hands :-/

Try to use this VAE: https://huggingface.co/stabilityai/sdxl-vae

@KandooAI  7.0 but since then i've been using 8.0 and it's good

@KandooAI Am working on a model for generating humans. It has done well at generating super photorealistic images. Now, I want to put those people in any setting, like in a kitchen, taxi rank, or bedroom, holding something, etc. I have collected all datasets that I want these people to be featured in. And have captioned them. Now am stuck as to how to organize that into training. Like

1. Do I throw them into one image folder or do I make separate folders for each concept.
2. Which training type do I go for - finetuning, lora or dreambooth
3. How do I deal with the issue of one concept being overtrained while others undertrain
4. Or do I train each concept separately and then merge all of them at the end

@sphiwe I don't know anything else, but I've done a lot of testing on the merge model. As far as your goals are concerned, you can rule out the #4 method. Because by doing so, you will also face time to merge tests > time to do training

After dozens of tests, I found that it's connected with amount of VRAM.

On 8 VRAM I can generate 512x512, but only if I run bat file with --lowvram.

If you have such error you need a major hardware upgrade to at least 16 VRAM.

This is not the place for technical questions. Even if it was, your question is low quality.

@Fooky9 lol genius

Use xpenis 1.4, 1.5 has problems

@rime11 I’ll try it, thanks!

@rime11 Works better with 1.4 but still now perfect. I will keep experimenting more.

@brownbagel0 Personally what I do is generate an image without the xpenis Lora and then inpaint the penis into the image with xpenis1.5, works perfectly for me

@brownbagel0 I think I keep the strength of the Lora at 0.5

@brownbagel0 just posted an image of what it would roughly look like

@rime11 inpaint with ?

@BNP222222 with xpenis 1.5, draw the penis and in the prompt put huge cock or something like that along with xpenis 1.5 lora

@BNP222222if you are asking what model i inpaint with I dont inpaint with any inpaint model but you can

v7 works for me in this context

@KandooAI at the risk of being annoying, I want to throw my vote in for a turbo model, too, lol. Still on an oooooold GTX and it's painful with SDXL.

@uncledude @crowav98995 You guys know that you can just merge XL turbo with Juggernaut and get Juggernaut Turbo?

I'm using this: martyn/safetensors-merge-supermario: Merge safetensor files using the technique described in "Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch" (github.com)

@KandooAI I also couldn’t create a photo with a full body, I’ll try as you said, thanks 

@KandooAI As I understand it, to make this work better, you need to describe the lower part of the body, shoes, pants, and so on

The more technical prompt for that is full length portrait. But the better way to achieve that is to put a full body picture in openpose, then leave only the body bone of the figure, and scale them down as small as you want small the depicted body. It will figure out many poses, even sitting or lying, but almost always with a full body. Set the controlnet weight lower to give it more freedom.

Same here, full body shot and full length portrait don't work for me

Same "issue" : Big difference with version 7 where the model could infer the whole body without having to give extra details (shoes or anything else): "full body" alone was enough and gave an image adapted to the context of the prompt. Version 8 is far less flexible.

use controlnet, google it

long torso is usually due to generating in a non standard resolution what image sizes are you getting long torso on?

Look like it is (768, 1344) also, have you noticed that the eyes are all pretty much the same shape?

Use "full body shot" in the prompt. Use "close-up" in negetive prompt.

Realistic background, detailed background

@anelesmelchiorre @GodOfSmething 
it doesn't work, I also try "bokeh" in negative message but no change, I think sdxl is the most trained in professional photography this is the cause of our blurry background...

@iphono that's unfortunate. Can you share the prompt and negative prompt and other parameters you use?

try 2 samplers

Not sure what link you are referring to, but that sounds like you're using a VAE, and maybe one for an SD model instead of SDXL. In this case, Juggernaut has the VAE baked in, so you should set it to None.

@Manax Ohhh perfect, yes i had selected some SD VAE, thanks

dwonload the picture and extract png info, all the setting is there

same problem on 3060 12gb cant get same image

Hey, you need to be using the exact same generation info - the listed configs don't show everything, click the copy button on the bottom right and then paste these details somewhere. The hires stuff isn't as important but the image dimensions etc are.

Your running on a local pc, what do you expect , this site runs off a server,, its gonna be different

Automatic1111, ComfyUI and so on and so on .. All those UIs implement different elements their own way, and one of those, the Sampling, has a huge impact. To make it simple the sampler ('Euler', 'DPM', 'DDIM' etc..) gives the schedule and the way the denoising is done in the latent space.

If the code below those samplers can be similar, some UIs will use the CPU and some will use the GPU (ComfyUI) for the scheduling.
So in the end unless you are using the same UI, your results will be quite inconsistent with someone using another UI.

I had some strange issues with Searge workflow. Never had any issues with juggernaut and my own workflows

Of course it works. Did you try the right control net for SDXL? the 1.5 doesn't work for SDXL.

@diogod that was the problem, thank you

Use "bokeh, depth of field, portrait" in the negetive prompt. Also, look for the name of a camera or photographer that matches your desired dof, and use the name in the prompt. Finally, yiu can look for an already generated image that matches your desired outcome and start remixing it, until you get what you want.

@GodOfSmething so kind of you i ll try and see what is gonna happen.Thx~

@cyptoRabbit anytime!

Hi my new reales Lora will help you with that issue to, Its called "Add UHD Details & Text XL and SD" hopefully that will help you out

@DiceAiDevelopment Thx! i ll try it right now

https://civitai.com/user/DiceAiDevelopment

You can find users by changing the search dropdown (left of the entry) to Users and then searching for them name

Oh My Days just noticed my name mentioned and was like wtf lol

Yes.

As a rule of thumb, we should generally use sdxl "extentions" with sdxl models.

@GodOfSmething many models are trained on post-refiner photos, and using refiner with them lowers quality

no that would lower the quality and is more resource intensive.
you can test a few photos with and without the refiner and see for yourself

You don't need LORAs to create realistic pictures with this model. It will react well to normal prompting like 'A photo of ' add some lights, some background ideas and the model should react well.
Try the 'Fooocus Photograph' or the 'SAI photographic', with the normal 'Fooocus V2" and Fooocus Enhance' and 'Fooocus Masterpiece'.

In the bottom of the 'Settings panel', you get a 'History Log' which will give you a nice tabular view of the images you have created during the day with the technical data associated. That will help you to see what works well or not.

You only need a Lora if you want to create a specific type of character or certain type of clothing. Like if you want to make a picture of a Disney princess, you could find a Cinderella lora or something.

I can't run it on Draw things cause I'm a noob, can you make a tutorial for that?

@mattiacusani665 Download the file from here and go to the drop down menu under "Model" in Draw Things. Click "Manage" and a window with all the models will show. At the bottom of there are three options, click the middle one (the import symbol). Click on the first drop down menu under Downloaded File and then select from files. Choose the file you downloaded from here, change the Fine-tuned image size to 1024x1024 because it is an XL model and then import the custom model. Have fun!

@Kind_Grapefruit_2351234 I've done it with default options putting it but grey results

this is my config: {"targetImageWidth":1024,"seed":2601398831,"numFrames":14,"hiresFix":false,"id":0,"controls":[],"hiresFixStrength":0.69999998807907104,"refinerStart":0.85000002384185791,"clipSkip":1,"steps":30,"imagePriorSteps":5,"hiresFixWidth":768,"hiresFixHeight":768,"maskBlur":1.5,"negativeOriginalImageWidth":512,"negativeAestheticScore":2.5,"seedMode":2,"sampler":0,"negativeOriginalImageHeight":512,"originalImageWidth":1024,"targetImageHeight":1024,"zeroNegativePrompt":false,"batchCount":1,"motionScale":127,"negativePromptForImagePrior":true,"height":1024,"model":"juggernautxl_v8rundiffusion_f16.ckpt","imageGuidanceScale":1.5,"cropTop":0,"clipWeight":1,"aestheticScore":6,"originalImageHeight":1024,"batchSize":1,"cropLeft":0,"fps":5,"loras":[],"guidingFrameNoise":0.019999999552965164,"strength":1,"startFrameGuidance":1,"guidanceScale":7.5,"width":1024}

In my opinion V7 best so far idk what you did guys after that but V7 was good excellent

@KandooAI take your time polishing the next update, we're rooting for you!

Seriously .. you have already mastered this version and extracted all you could from it?

@KandooAI Thank you, it helped!

@KandooAI I removed 'vae' and the result changed. Now I'm getting not just a bunch of pixels, but simply bad images, even though I follow prompts and examples

@figusqw180  that linked image was the literal output? 512x512 ... change the output size to 1024x1024,,, XL checkpoints generally do not output useful at 512

Make sure to merge in the VAE. Use the SDXL fp16 fix VAE (google it) or jump in the RunDiffusion discord and we’ll link it for you.