I'm not into strong specular highlights, so that's good to hear. "Colors look flat"? Flat like a regular TV anime?

@kintilian45 yes

I would have to agree because, apparently the thirteen version would not Let me transform the body form of shape, I like.

Yeah, after trying v13 for a while I switched back to 12.

V13 seemed to give me more artifacts, and was harder to get sfw images in some cases.

WAI is still my go to model though. I train on base Illustrious, but generate with WAI.

They all seem to produce different-looking results. I'm glad we have the options, but I'm not looking forward to losing access to v10-12 on the generator when newer versions come out

0.1 i think

I use ComfyUI, and there's no doubt it can be used in ComfyUI — the method is the same as with SDXL models.

Yes, you can use this with comfyUI.

Assuming you have comfyUI setup and ready to go.

You can download the checkpoint from the top of this page.
Put it inside your ComfyUI/models/checkpoints folder.

Startup comfyUI (if it’s already open you’ll need to refresh the page).

There should then be a Workflow tab at the top that you can click on. Then go to Browse Templates. I think an Image Generation option should be one of the first things you see. So, load that one up (just ignore a msg if it says you’re missing a model).

You should see a Load Checkpoint node; you want to change the option there to the checkpoint you just downloaded

Then you hit run or ctrl + enter to start the generation.

That should be the basics to get you started.

@LightSwitchy Thanks for your answer and time!

V7 was also great. After that, it became more anime-ish

Can you be more specific about what got worse?

I was wondering what version would be best. Appreciate the heads up, thanks.

you have to put it in the Stable-diffusion in the model folder

@Theunknow106 Already. When load, it said Error, can not load models

i run it with directml works fine for me, try use this versions with your build, try em out one at a time see if they work for you, diffuser=0.30.0 and also, transformers=4.25.1, another thing, you shouyld try using the latest AUTOMATIC1111 build v1.10.1, they resolve a lot of issues with directml, for me it worked instantly just changing those to versions in the requirements_versions.txt file

thank you !!!! awesome breakdown !!!

What is the principle of choosing the first pass？

First pass applies the refiner to the initial generation. Second pass applies it to the hires fix portion of the generation

How do I use it? When I use more than one checkpoint during generation, my GPU memory decreases and I get a warning. Is there any way to switch checkpoints without losing memory?

@inside1589 If you use forge, it should unload the first model before loading the next one. I can't remember if A1111 does the same

I've used refiner with fooocus, but I've no idea how to make workflow with comfyui

https://followfoxai.substack.com/p/part-3-sdxl-in-comfyui-from-scratch should do maybe

For me the eyes usually get fixed by just using ADetailer. Thx for the advice anyway, will try it out.

i gotta say illustrious 1.1 is not so good tbh, maybe its just me but that fine tunning has the same problem as ANIMAGINE XL v4, its somehow not that good

I heard there are issues with Illus 1.1 due to it being trained with a sizable chunk of AI images and overfitting

@rivertree same with animagine v4 ive tried everything but still no good results, seems animagine v3.1 is the best still !!!!

illustrious v1.0 is already refined enough to go crazy for next the 5 years !!!

Illustrious 2.0 just got released and i hope they'd use that instead of 1.1 if they can

@tobycortes WAI is using illustrious 0.1 no? not 1.0 from what ik

@DawnDenn It would be crazy if WAI wasn't updated to this even after the 3.5 release.

They definitely do for me.

Try artists with more than 50 danbooru hits, ideally 100+

if they have brackets in their name, put a backslash before each bracket.

Use spaces instead of underscores, unless the underscore is an actual pat of the artist name.

DO
squeezable \(squeezabledraws\)
DONT

squeezable_(squeezabledraws)

its does work with the correct syntaxis as stated above

@FondantViper does this apply to other prompt as well? Because most tags on danbooru has underscore (character name, anime name, etc.)

@Kaze111 Yeah. If you're using a character name or series name, same thing applies.

DO
misty \(pokemon\)
DONT
misty_(pokemon)

it can get pretty crazy for long tags. One of my favourite outfits is very long to correct so I keep a text file for easy copy pasting.

artoria pendragon \(lancer alter\) \(royal icing\) \(fate\) \(cosplay\)

1. What resolution are you using?
2. What sample and scheduler?
3. step count
4. what GPU?

@FondantViper 1. 680x800
2. Euler a & Automatic

3. 90
4. I got a 4060

@Von_BlepThe3rd do 30 steps not 90

I haven't tried yet, but perhaps this will help:

https://education.civitai.com/civitai-guide-to-controlnet/

Also, any search engine: "Controlnet: A complete guide" pulls up lots of other-site results.

@croutucreicesoi657 those dont seem to work for illustrous, I understand nothing

Do you use Comfy of stable diffusion?
If you use comfy then there are nodes that extract a skeleton from a reference image. (same with midas)

For me the structure is like this: load image>pixel perfect resolution>get image size>AIO aux preprocessor> preview image. In the preview (or save image) you'll get the skeleton or depth map you want.

Keep in mind not all can be used for a skeleton, some simply don't work.

You can use either the Openpose preprocessor or the illustrious one uploaded to this site.
If you use Stablediffusion I sadly can't help you.

@Hati_Hrodvitnisson I made it work, thanks. I have other questions if you don't mind. How can I make a character wear an outfit? Like for example I made my own oc, but I want her to wear rem maid outfit, but I can't find how, all I get when using rem cosplay or roswaal mansion maid uniform is making the girl look in the re:zero style, do you can help¿

@almaqueso Basically what you have to do when using Illustrious or Noobai is make use of danbooru, there you can find tags that suit your needs.
In this instance for example you prompt your oc and then you can add rem \(re:zero\) \(cosplay\).
you could also add it beforehand depending on which one yields the best results.
But alot of prompting is knowing the right words (danbooru tags in this case) and structure.

It depends on what you want , but from personal experience:

- Use SDXL 1.0 as training base if you want general purpose compatibility across many SDXL-based checkpoints, Since is the official base, your trained model should work with every other fine-tuned SDXL checkpoint and not just Illustrious-XL based fine-tunes like WAI-NSFW (this checkpoint) or NTR Mix.

- Use the base fine-tune (Illustrious-XL) for training if you want wider compatibility for you model with most checkpoints that are also based on that specific fine-tune, such as WAI-NSFW or NTR Mix. This will also ensure that your model works well with future versions of WAI-NSFW.

- Use the specific fine-tuned checkpoint (WAI-NSFW, NTR Mix, etc) as training base if you only want to use your model on either of those specific checkpoints but doing so may make the model not compatible or difficult to use for any other fine-tuned checkpoint that's not the one that was used for training, since your model would become biased toward that base fine-tune aesthetics. For Example: a LoRA trained on WAI-NSFW v13 should work well for that same checkpoint but may struggle to generate good results on NTR Mix or even future versions of WAI-NSFW.

go with noobai. this model was trained on ai slop images

@HentAI_Ga_Kill with all due respect. if you have such strong opinions please post some pictures on noobai so we know the difference between ai slop trained model vs the good model.

Thanks for comment guys. I've narrowed it down to Noob and ILXL 1.x

works

it's worked, but you need to adjust many (and I mean a lot) of things if you r using prompt from the former version

這個應該放ComfyUI\models\checkpoints裡面

diffusion_models是用来放flux模型的，你搞错了

same problem

好像没有，v13是2.0出来前发布的

@1344535564102 好的

I think v13 is a slight downgrade.

I find v12 to be better.

v12 is better than v13

832x1216
1216x832

1024x1024

@Nephilim ok thx

i have ,i not know how to avoid it ,its like star,but i dont have the prompts like star,i dont know why

@zpmljzzc I might have found a solution. Setting the Denoise option on the second sampler to 1.0 seems to fix it

@Tuti03 thank you,i will try.i found another way ,i put the words"particle,light_particles, particles, "in the negtive prompts，i found it used,please try it

i use comfyui

I think it's either that impurities are included in the training dataset, or that the color tone settings incorporated into the model itself are incorrect. This was common during the SD1.5 era. If the latter is the cause, it can be fixed using SuperMerger's adjust function.

@watayuki288 thank you very much! one question a little,i use comfyui,i look github,buy not find the nodes about Supermerger,thank you

@zpmljzzc use a different vae

@ofnilthefinegame thank you.i use some vae,but a point a point in the picture also,for example,i use black_backgroud,the point like "star",but i not use the star in prompts.

@zpmljzzc Supermerger is a WebUI exclusive extension. ComfyUI is specialized for image generation, so I think there are no extensions for model modification purposes...

@watayuki288 thank you!

i usually put things like "watermark, signature, text, bad quality,worst quality, patreon signature" on the negative prompt and it seems to work fine.

put (text:1.5), (watermark:1.4), in negative prompt. include the brackets.

As the others have commented, you need tags in your negative prompt to reduce the chances of watermarks, however V13 of this model does seem to force watermarks a bit more than previous versions, regardless of what you place in your prompts.

@dot_original @Cardin @MaGiC48x im using these negative prompts but im still getting them, could it be that certain loras make them appear more often?

@yoyo12333 it would greatly help if you uploaded examples so we can analyze, meta-data important so make sure it's included. on site generations will have it.

generally it would be the checkpoint's fault as the checkpoint is trained on a lot of images with logo/watermark. the negative prompts should cover it however, you may be on a streak of type of images that are very rigid or something. hard to tell.

if you are afraid that your images are to explicit to upload or so, scrap those thoughts. you are anonymous here and there is just no way you will have more explicit images than what is already on site. you can also just delete them later.

@yoyo12333 you're not doing anything wrong and I doubt it's your loras. V13 of this model just seems to force watermarks more often than previously

AI is similar to 3D, the higher resolution the higher the quality as the rendering 'software' has more pixels 'to work with', problem is with AI you are bound to find more body deformation the higher the resolution, newer models like Noob even claim to be able to generate at 1824p without the so called 'body horror'.

I do the same, generate at 1024~1536x1024~1536 then upscaled by 2 at a denoise of 0.34 to make it as close as possible to the original image but reducing or removing artifacts entirely thanks to R-ESRGAN 4x+ Anime6B (the best anime upscaler)

But keep in mind, my workflow is agressive, pure gacha, there is nothing wrong with it, but most prefer 1.5 to avoid changing the image to much, I on the other value sharpness and fine details, so I got with 2 and 34.

As for your Denoise think like this, 0 denoise the image is the same as the original image, at 1 the image is entirely different than the original image, you have to find a value that works best for YOU.

@xDegenerate is 100% correct with their comment.

I'll just add, for me personally locally generating on an RTX 3090 in Forge, this model seem to work fairly well on initial resolutions up to and including 1408x1408. I then hires fix upscale to 1920x1920.

From a lot of experimentation, resolutions any higher than this seem to drastically increase the chances of body deformities and graphical errors.

Stable Diffusion XL is trained and optimized for 1024 x 1024, that is why this will give you the best image quality. You can do it higher, i also got good looking results without body horror with generating directly 2048 x 2048 with some checkpoints but as SDXL is optimized and trained on 1024 x 1024 higher resolutions should actually decrease the quality. So i use only 1024 x 1024 for square pictures with my RTX 4090 and then use Hires.fix with R-ESRGAN 4x+ Anime6B upscaler and upscale by 2. But you can also get good results with higher initial size. and for vertical format i use 832 × 1216 or 1024 × 1536 also upscale by 2 with Hires.fix.

Какое нафиг "норм", когда оно всирает стили и глючит по сравнению с двенашкой!..

@Ptiz уга-буга

Yes, 'anime screencap' will desaturate the colors while adding some form of 'gaussian blur' to it, making the render less sharp, and 'anime coloring' will make the lines more thick and will have some saturation compared to 'anime screencap'.

Use both if you want to replicate anime styles (duh).

If you don't use any of these : Anime Coloring, Anime Screencap, cartoon or Toon (/style/) your renders will have by default a more soft-shaded look unless your ckpt is specially trained for anime like Mistoon.

Most people just add tags they don't know how it works, they just mindlessly copy from other users.

@xDegenerate Thanks for the info. I have a potato + it can literally take all day to test simple tweaks and learn how they work. I myself was a mindless copier until learning to optimize properly / after a couple months at it...now I'm only a couple years behind :)

@xDegenerate interesting, thanks for the info.

I been mostly using pony where these tags certainly helped with lora trained on screencaps for newer shows or limited datasets, noticed i was getting pretty bad results until i removed them, sounds like it was probably the screencap causing the issue, i will have some more playing with them, the default (not using them) certainly seems pretty nice. I admit i avoided Illustrious for some time because i have limited time and didn't wish to learn new system when busy but it's actually pretty fun so your input is much appreciated as i play around.

It looks IL 1.0 now 14V. The reason why he didn't take 2.0 was that it didn't meet his expectations.

1024x1024 is the officially supported resolution.

I've done a lot of testing with every resolution past 1024x1024, increasing by 64x64 every time, and counting the amount of body deformities and graphical errors.

1024x1024 obviously has the least errors, however 1280x1280 and 1408x1408 also seem to work very well for me.

Added bonus of the higher resolutions is noticeably more detail, but the risk of body deformities does increase, especially at 1536x1536 and higher.

Should mention this is through local generation in Forge on an RTX3090

1216x832 is what I use. Upscale by 1.25 just to add detail.

Generate a 1024×1024 image first, then upscale it using Ultimate SD Upscale.

If you're concerned about detail loss or differences after upscaling, you can combine Canny and Color Transfer to help preserve the outlines and keep the color tones consistent.

Personally, I found v10 to be the best one when used with Loras. Tried 11 and 12, but both were worse with loras and also inconsistent with things like fingers. v11 was really good at regonizing and copying know artists compared to v10 tho.

You will get mixed answers here on CivitAI as some people can be biased (heavy simping), the first day v12 got released people were being downvoted to hell for speaking against the constant updates.

v13 was not much different but more people spoke against it.

The creator excuse as to why he was updating frequently was because the ckpt was generating to many 'minors'.

My opnion is that he was simply farming for likes/buzz, after all this guy is the most well paid creator in the whole site (do the math on the Buzz assuming 1$ = 1K Buzz), plus he is making on average 1.5m Buzz week with this cpkt alone, so he is obviously trying to stay relevant making frequent updates.

As for my opnion on the v13? is uber dogshit, the anatomy is completly destroyed, zero support for futanari content either (as much some people dislike it, it is a thing), anything higher than 1280p and the body deformations are just to much.

v12 is OK.

The best imo are v10 and v9.

v9 is simply an all arounder, if you make a 0.33/0.33/0.33 merge with Personal Merge and Hassaku Style A you got a model that can meet basically all your needs (Minus 3D and Realistic).

@xDegenerate I thought it was, for some reason, my fault, either my generator, my own loras, or my prompts, but indeed with 13 (which I costantly use), give some weird deformities some times. I might go back to 10 and check if it happens as well. thank you for the insight :)

@Liviel Worse in what way? extra fingers etc?

@xDegenerate Wow, thanks for the insightful answer. I will try my Loras in both version because in V11.0 it's a bit unstable, but stilll good.

@Andrach Extra fingers, artifacts, deformed fingers, and when it came to loras they were worse at mimicking styles.

@Andrach arms becoming body and viceversa, legs reversed from front to back, etc lol

its related to VAE in your case since you mention looks great in preview but came out ugly last second, usually the fix is increase the steps from 20 to 25 and set the VAE to auto

Please use a version prior to V13, as this issue only occurs in V13.

I haven't encountered this issue in other versions, but in V13 it happens much more frequently.

@alona_chemerys Thanks for the help, i changed VAE to automatic and now it works great

I think its 0.1

@ps1copato :(