Danbooru

Pokemon6242  Oh, thanks, I thought it was in the tag groups, so I didn't even pay attention to the top index. I'm a bit dumb, so I really appreciate it.

V15 wont be released for awhile. The model creator replied in another comment that there isn't much room for improvement with V14 at this stage and is going to wait for the Illustrious V3 base model to be publicly released for download.

A WAI model based on Illustrious V3 would be incredible with all the features it has, however looking at the Illustrious website, we're probably still a decent wait from an open source release of Illustrious V3.

MaGiC48x yup, gonna be 3 months minimum. Realistically 3-6

ZeroHour1 Illustrious want ten times the amount of funding for a V3 release compared to V2. (V2 was 1.5million stardust. They want 15 million stardust for V3 release).

I genuinely think we'll get Grand Theft Auto 6 before Illustrious V3.

MaGiC48x Where's that 1.5 million stardust coming from on the one I'm looking at it says 300,000? Unless I'm looking at the wrong thing.

DatCivitAIGuy 300,000 is for a closed source release for contributors.

You need to look at the "Global Stardust Progress Meter" here:
https://www.illustrious-xl.ai/sponsor

Illustrious have been a little scummy and misleading with their fundraising by having the two separate Stardust goals for each model.

There was a guy that paid the full 300,000 stardust amount for Illustrious vpred v3.5, thinking it would mean an open source release. They took his money and only released it closed source, meaning we still can't use it. Illustrious vpred v3.5 needs a ludicrous 53 million stardust for open source release.

This mess has pissed off a lot of contributors, hence why I don't think we'll see Illustrious V3 anytime soon.

FYI $1USD = 100 stardust.

MaGiC48x That is scummy as hell. Do they really think they're gonna make over $150,000USD within a short amount of time? Like holy crap they're on something. Good thing you told me I was actually about to make that same mistake and waste $1,500 USD to push it to the final goal. Well I guess that's a no go.

DatCivitAIGuy I fully understand that training AI base models isn't cheap, hence why Illustrious want crowd funding, however the huge jump in price from V2 to V3 is too large, especially considering how lackluster V2's release was, as well as strong competition from other models like FLUX and WAN, that are free to use.

I'm pretty sure this model already has a VAE integrated into it so you don't need to use that VAE.

obsm0487389 i'll give it a shot thx

yh

Do you use the Moriimee Style lora at high strength ?

DrLox Yes I do, is that the problem? If so, how do I fix it?

Average_Gooner It's an issue with the moriimee's lora, so you have to lower a bit the strength and to add "doorframe, door, frame, doorway, corridor," in your negatives.

realised I had my node set up wrong, instantly worked once i fixed it!

what is Illustrious 2.0? has it launched?

He said that he's skipping IL 2.0 as he's not satisfied with the quality and waits for IL 3.0 public release

Use eyeful_v2_paired.pt.

I use ADetailer with faceyolov9c and get perfect eyes 99% of the time

These are pictures that they generated locally on their own GPUs.

since it shows how much images are generated by Civitai's generation service. It doesn't count people who generated it locally on their desktop

Swiney I just tried posting a locally generated image (https://civitai.com/images/94391760), and it appeared in this model's gallery. So i'm right.

empanada_48 aren't we saying same thing? locally generated images are actively uploaded in this page, but temperature system only cares about image generated in-site, so page says 'dormant'

Swiney You are, bro can't read.

Swiney I re-read your initial sentence. You're right: you meant that the term ‘Dormant’ refers specifically to generation on civitai, not locally generated images. Well, shit happens, my bad.

v9 and v10 is the most stable one for me.

v9 - the best

V13 is actually the most up-to-date, knowledge wise. V13 has some visual issues though so you'll need to use it with a refiner model too. I recommend V11 as a refiner.

@MaGiC48x Thanks!
I'll probably stick with V12 and use V13 with a refiner model when needed.

are you imputing the activating prompt of the style lora?

ArtistIA Yes, both the trigger word (if that's what you're referring to) and the actual lora input, both at the strengths in the example images and recommendations on the main lora page.

Dstack thats weird...

a quite annoying quirk of Forge is that if it can't find a lora it won't give you any error. or warning on the web page. You get something if you check the window where you launched the forge server but it won't tell you what lora is missing either...

I trained a dedicated LORA for the character’s eyes and used it in FaceDetailer. However, if the training data for the character’s eyes in the LORA is insufficient or too small, details are likely to be lost.

@koier987877 Thanks for your answer. The problem is previously when using this checkpoint I wasnt getting trippy FX with the eyes. It only started happening this month and for some reason the online generator just continues to generate these various trippy eyes, when last month or previous month it had no problem generating normal plain eyes.

Have you tried using other versions of WAI-NSFW-illustrious-SDXL to see if the same issue occurs?
Also, try giving your entire prompt to ChatGPT for analysis—it might help!

@koier987877 Currently using version 14 and not getting trippy eyes because I borrowed someone elses seed, it might be related to which seeds you use, the eyes are not trippy with any FX, although they arent exactly perfectly detailed (a little blurred) and would need hi-res fix or something, but I am not going to worry about it for now. I will let you know the link when I post the images.

@koier987877 check the link in 25 minutes https://civitai.com/posts/21625619

I tried generating an image using the same LORA configuration you used for testing. Here is the image:https://civitai.com/images/97521280
Using Embedding can help improve image generation quality and prevent certain errors. I checked the LORA you used, and the images generated by other users had normal eyes, so this is not a problem with the LORA.

Use this Embedding:
https://civitai.com/models/1302719/lazy-embeddings-for-all-illustrious-noobai-pony-sdxl-models-lazypositive-lazynegative-positive-and-negative-plus-more?modelVersionId=1558647
Use the lazympos Embedding in the positive prompt, add lazyneg to the negative prompt, and then try generating again.

You can also use Civitai’s Image-to-Image Face fix feature to correct the eyes. When doing so, keep the eye-related prompts, art style, quality LORA, and quality prompts, and set the Denoise value between 0.55–0.7 for the correction.

@koier987877 Thanks for the info. I can't apply some of the info because have a free account and embeddings are no longer available for free account users (they where in the past). Also if you use the online generator, it has a memory and sometimes it takes a while to flush out your old prompt (sometimes previous prompts you where using will continue for more image generations until it gets flushed out eventually, or sometimes some of your prompts could be delayed in getting generated and get generated after a few attempts and I have experienced a long delay lag in some prompts getting generated, then get generated when you have removed the prompts being affected, so there is some sort of memory system with the online generator). Other times, (I cannot prove this but it is my belief through observation) if the online generator is glitching, images can be generated with other user's prompts, I remember for example one day my characters kept winking, without any prompts used for winking and without any specialized resources used for winking. That's when the memory in the generator has glitched. These are my opinions.

When the seed is random, the model doesn’t actually retain memory, but internally it has already assigned higher weights to certain prompts, which leads to the desired result.

AI models tend to hallucinate randomly in areas you don’t explicitly specify, which is why embeddings or LoRA are needed to greatly reduce those hallucinations. I’m very strict when it comes to eyes, so I always put a lot of focus on that part. Basically, if you want to check whether a model can draw eyes properly, you should use the facedetailer from ComfyUI-Impact-Pack to do an eye upscaling fix and then look at the result. If the fix still looks terrible, then it’s likely a model issue. I also use the same method to test whether a character LoRA has properly trained the eyes. I’ve been using the WAI V14 model for a long time, so I’m certain that its eye rendering is fine. When problems appear, applying upscaling fixes is the right solution. In general, the smaller an object is relative to the whole image, the more likely the AI model will generate it incorrectly.

Unfortunately, if you can’t use those features, you’ll just have to rely on luck and hope for a good result.

@koier987877 Thanks for your info again. Just a query, whether you ever use the online generator or mostly use local setup such us Comfy? If you dont use this online generator (i.e generate 100 images in 1 session) then you wont really understand some of the memory anomalies that can be experienced. Is it possible the online generator itself has a memory system that can sometimes interfere with generation results, is what I am asking?

@koier987877 quote 'eye rendering is fine'. I agree but for at least a day I was getting those hypnotized spiral patterns in the eyes and it would not go away no matter what prompts I tried and I believe someone else using the online generator that day generated those eyes for their character and somehow their prompts poluted my prompts through a glitch that day. ???

@koier987877 quote: When the seed is random, the model doesn’t actually retain memory,

I thought if you change the seed number by 1 increment the next image generated will have all the memory of the previous image but slightly vary the image because the seed variation was absolute minimal. Naturally the spiral eyes anomaly will continue but I believe I increased the seeds dramatically and may have even changed the seed to a random number but still was receiving spiral eyes. Normally when the online generator continues generating some anomoly the only thing you can do is reset the generator but you also need to log out of account then log back in because only then will the anomaly be flushed out 100%

I mainly use ComfyUI to generate AI images, because ComfyUI offers a lot of customization options and has many useful plugins.

You might think that AI can render perfect eyes, but in reality, AI models are not that perfect.

You don’t use tools to fix the image.

you doubt that the online generator might have some kind of interference. That mindset isn’t good — it will limit your progress.

If you truly want to create the perfect image you have in mind, you first need to understand the tools and techniques you can use.

After trying and confirming that it really doesn’t work, then use another method to achieve your goal.

Remember that the AI tools or models you are using now are not that perfect.

 quote: When the seed is random, the model doesn’t actually retain memory,

You need to understand that the images generated by an AI model are all “calculated.” Therefore, if you use the same model, the same prompt, the same seed, and the same parameters, you will get the same result.

While the seed does affect the calculation results, it is mainly the prompt and the chosen sampler that influence the outcome.

But why?

usually they release 1 version per month,but the lastest(15) was in may

PlumpBouncing I see. But do they do it because of their framework or just for the sake of keeping updated, because so far the model is almost perfect.

SAAIKOPATH don't take what i said too seriously,i'm just kidding because they don't released a new model,but i understand what you said ;) ty

PlumpBouncing Nah, it's not a big deal. I was genuinely curious to learn if they had a certain way of doing things.

WAI has stated V15 won't be released until Illustrious V3 base model is open sourced.

@MaGiC48x illustrious v3 is not going to be open source for a long run. i keep updates on illustrious website can those money hungry people reset the donation goal ever time.

It is recommended to add the following prompts:
,(uniform lighting, soft shadows, balanced exposure, even illumination, smooth shading,)
,((matte skin)),

v12 is my go-to, v14 has way worse anatomy

i just stick with v7 or v8. other version feel like can't accurately make character clothes.

v11 looks like semirealism

If you generate locally, you can use two versions at the same time. One version as the initial model, and the second as a refiner. I personally use V13 as my initial as it has excellent prompt adherence, but it doesn't look great, so I use V11 as my refiner to fix all the visual issues that V13 has.

If you haven't used the refiner, the way it works is you'll provide a percentage that the refiner will switch, then it'll automatically swap models at that percentage of total steps.

Example:

50 total steps, with refiner switch at 40%.

Model V13 will generate the first 40% (20 steps), then it'll switch to V11 as the refiner, and complete the remaining 60% (30steps) using just V11.

WAI stated in another comment awhile ago that they are waiting for Illustrious V3 base model to be released open source, before starting V15.

WAI mentioned there isn't much that can be improved over V14, at this stage.

Unfortunately, we're likely a long way away from an Illustrious V3 release due to the sheer amount of funding the Illustrious team are asking for, as well as their misleading crowd funding campaign, which has deterred people from donating.

@MaGiC48x Thanks for your answer :)

I recommend setting clip_vision to CLIP-ViT-bigG-14-laion2B-39B-b160k (size 3.69GB), and using NOOB-IPA-MARK1 for the IPAdapter.

https://civitai.com/models/1000401/noob-ipa-mark1

You can try the following parameters
clip_name: CLIP-ViT-bigG-14-laion2B-39B-b160k

ipadapter_file: nooblIPAMARK1_mark1.safetensors

end_at: 1.000

sharpening: 0.20

embeds_scaling: K+mean(V) w/ C penalty

weight: 1.00

weight_type: strong style transfer

combine_embeds: concat

start_at: 0.000

If nooblIPAMARK1_mark1.safetensors does not work, please re-download CLIP-ViT-bigG-14-laion2B-39B-b160k and make sure its size is 3.69GB.
CLIP-ViT-bigG-14-laion2B-39B-b160k
https://hf-mirror.com/h94/IP-Adapter/blob/main/sdxl_models/image_encoder/model.safetensors

DPM++ 3M SDE with Karras

i have tried both and it doesnt seem like it matters, i copy/paste artist names a lot and it understands them with or without

From my own tries, adding and removing underscores seems to change the images drastically, and it's probably better for you to stay without underscores

I’ve been using WAI V14 together with IPAdapter, and it works properly.

I recommend setting clip_vision to CLIP-ViT-bigG-14-laion2B-39B-b160k (size 3.69GB), and using NOOB-IPA-MARK1 for the IPAdapter.

https://civitai.com/models/1000401/noob-ipa-mark1

You can try the following parameters
clip_name: CLIP-ViT-bigG-14-laion2B-39B-b160k

ipadapter_file: nooblIPAMARK1_mark1.safetensors

end_at: 1.000

sharpening: 0.20

embeds_scaling: K+mean(V) w/ C penalty

weight: 1.00

weight_type: strong style transfer

combine_embeds: concat

start_at: 0.000

If nooblIPAMARK1_mark1.safetensors does not work, please re-download CLIP-ViT-bigG-14-laion2B-39B-b160k and make sure its size is 3.69GB.
CLIP-ViT-bigG-14-laion2B-39B-b160k
https://hf-mirror.com/h94/IP-Adapter/blob/main/sdxl_models/image_encoder/model.safetensors

You can merge it with other models, if that suits you, but you don't need to.

It is a complete model in itself. I recommend using it alone, or any lora you prefer.

not happening for few more months

do you use any artist style to make anime like images?

... the usual way, like any other SDLX-derived model. Simply load this model (and remove the refiner part of the flow if it's there, it's usually not required)

yes

in negative

First, run the image through WD14 Tagger to identify the corresponding tags/prompts, then add them to the negative prompt and increase their weighting.

For accurate testing, please remove the LORA during the test, as sometimes the issue may be caused by the LORA itself.

WD14 Tagger
https://github.com/pythongosssss/ComfyUI-WD14-Tagger

@koier987877 thanks, this is helpful. and to think it extended my negative prompt list by around 10 words to effectively reduce those occurences...

Impotence is something to be ashamed of, not proud of.

I'm sorry for you

Why would you advertise that on a place as porn-heavy as Civitai?

real

Then don't look at it? This is like going into McDonald's and telling the cashier "I don't like McDonald's"

Yes, bro, I don't agree with you, but I am surprised that you have the courage to say it.

6