they actually work together with the image generation web.

Don't worry, it will be open source in time!

For more information, please read the producer's blog

In reference to sigu_hk390, here is Onoma's plan for their releases https://onomaai.com/40

@sigu_hk390 1.1 is their finetune and it will not be opensourced.

@Volnovik "source: i made it up"

@nubby From official civitai newsletter: Version 1.1 Coming Next Week – A refined V1.1 will drop soon, but will be limited to on-site generation only.

Provide other sources. This week we saw that they can change their mind quite quickly

I also find it unethical to make us pay for stolen art

@Volnovik https://onomaai.com/40

"We’ll be rolling out our latest models step by step while progressively open-sourcing previous versions"

@low_channel_1503 It's still really shady to not have this in the description of the model.. like, why not just explain that? Why do people have to go and find it on a different site?? It creates unneeded drama.

Edit; I appreciate you linking where it was stated though

@low_channel_1503 we'll see. If it will be open in half a year - it doesn't make it open source now. Also they state for example, that 1.0 is nlp capable. In their technical paper it is stated that only 1.1 was trained with nlp tags

@Light7799 i agree with you. not sure why this isnt being announced

Latest version of illustrious is v3.5. So, we will see this joke again, again and again for v2.0, v3.0 and v3.5.

@Light7799  Tbf maybe it's a language miscommunication problem? If English isn't their native, I'll give them the benefit of the doubt with this one.

I still don't like the idea of paid models though.

@Ainme_Diffusion Even if I don't like it, I don't think asking for $10 for each model is that much of a problem. As long as the promise is kept about releasing models for download. They should just be transparent about it

@low_channel_1503 Investment for Illustrious XL is a good thing. However, the authors of NoobAI XL, Pony Diffusion XL, and Animagine XL did not demand any financial compensation for the release of their models. They shared their "latest" model to the community without any conditions and incorporated feedback from numerous users into their model updates. This is in stark contrast to the attitude of the Illustrious XL authors and makes us realize the true meaning of open source.

angelbottomless posted that on reddit.

The issue is that civit stated that 1.1 will not be downloadable. Also since backwards compatibility is not really working that well, I do not see any point in retraining loras for it. When 2.0 will be out? In half a year? or tomorrow?

@Volnovik You've got a great point, what is the timing between these versions? Are they just sitting on them and creating an artificial wait between models? Who knows.

Are for 1.1 being downloadable or not, that's on Onoma, not Civitai. Onoma has effectively declared that 1.1 is generate-only on all platforms it's on right now, with it going either free or paid download once the next version is coming out.

@cyniclandscapes well, I'm not blaming civitai for locking new toys, thats the only official statement we have about it. I am currently testing Illu1.0 (will have to make an article) and do not see much new it brings to the table.

This makes no sense and will end up killing/slowing down Illustrious progress. Why would anybody start merging, fine-tuning or training on a version that may very well be old by next week. Just release the newest model right away and get the community up-to-date! I highly doubt 1.0/1.1 or future versions of the model will be 100% compatible with tools and LORAs trained on, for example, v0.1. Current tools for Illustrious will very likely require re-training.

I experienced the similar compatibility issue. Many v0.1 LoRAs I had trained were not worked well on v1.0. The advertisement related to compatibility is not true.

Trash model no matter what. Quality on par with vqgan with prompt adherence worst than dementia.

"because it is necessary to get user profiles for model improvement" Could you explain how exactly this works? I'm honestly curious to know the steps and methods used. Like are they planning to fine-tune it with the gens made? Do they carefully analyze the gens? I'm interested to know

@Light7799 Nobody knows what exactly they plan to do with this, it's just obvious that they most likely would use outputs for training or something else - their terms of use state as much.

@munchkin I see.. just I remember Midjourney did the same thing, in that they needed people to use it and generate with it to 'train it' and they never really explained the process either. So was just genuinely curious of how it works

They also said so that; We will also share the insights from data analysis.
This will be beneficial to the civitai community.

People happily pay for NAI which is exactly the same, aswell as other services where people basically reupload pirated content. So yeah, it does work and will always work lmao. Problem is just how you do it.

@upscaleanon537 Did 1.0 get leaked or did it not?

@tombofamayan You mean NAI V1 or Illustrious 1.0? Because if illustrious, I have no idea.

@tombofamayan @upscaleanon537 Yes, seven days ago.  

Although, computing power and organizing data sets require costs, and adding the VPRED algorithm requires additional effort. The problem is not the fees, but how they are charged.

He launched a paid early version of 1.0, but released 1.1 on another POST in just 7 days,And it’s only available on online platforms.

lmao, were midjorney or dall-e leaked? did they compensate anyone? don't be so delusional, internet warrior

@dedededestruction I don't understand, you asked me if there was a leak and then proceeded to answer your own question. With a character attack no less, what are you trying to find out, internet warrior?

lol

This sequential release of new models isn't about catching up. What I see is the ONOMAAI team's architectural refinements. By strengthening the model's training afterward, it could surpass NoobXL. Besides, shouldn't a model's quality ultimately be judged by real-world performance data?

@liangren26 

1) OnomaAI has remained inexplicably silent for about four months following the release of ILXL v0.1, with no convincing justification for their actions. If their goal is profit-seeking or transitioning to a closed-source model, their current behavior is difficult to understand. (There may be internal company secrets at play.)

2) They tend to take an academic approach and favor novel methodologies. According to their research paper, multi-level captioning technology has been applied starting from the v2.0 model, suggesting potential improvements in natural language processing performance.

3) The NoobAI XL is an incredibly powerful model that leverages an extensive dataset and massive computational resources. Personally, I find its character database and art style reproduction capabilities superior to those of ILXL. On the other hand, the ILXL model excels in high-resolution output and natural language processing. While both have their strengths and weaknesses, I speculate that ILXL holds a slight overall advantage over NoobAI XL. (However, as of v1.1, the NoobAI XL model appears to be the better option...)

4) Since the latest ILXL model has not yet been released, a direct performance comparison is not currently possible. However, assuming there are no overfitting issues, I expect performance to improve with each successive version.

NovelAI has downgraded its current model so that its new version can work. It's also crap.

v1.0 will likely be available on civitai in a few days for download.

Originally, these guys completed the training up to 2.0 by September last year but are not releasing it. There’s no answer to this; if it continues like this, we might see a model that's better than Flux or SD 3.5. In that case, there would be no reason to use XL. Looking at the current situation, it seems that the release of 3.5 might take until May to July. It seems they have no intention of communicating, so it's best to just ignore them

In ILXL v1.1, the following characteristics were observed:

<High Resolution>

- Advantages: Increased character detail, enhanced aesthetic quality

- Disadvantages: Longer image generation time, unnatural poses, increased background noise

<Natural Language Processing>

- Advantages: Improved prompt comprehension, enhanced ability to distinguish multiple character traits

- Disadvantages: Reduced compatibility with the previous v0.1 LoRAs, potential forgetting of previously learned concepts with Danbooru tags

While I acknowledge the potential of the ILXL model, I believe its side effects are currently quite evident.

I look forward to future releases of improved ILXL models and various merged models that can help mitigate these drawbacks.

Noob can already do 1536x1536 though

@illyaeater no. its not native 1536. there is traces of the ai model trying to stretch up the art when using 1536. However its definitely true that vpred is a better prompt adherence model.

@Madafada1991 I mean you can say it's not natively 1536 whatever that means, but the images look completely serviceable. Just try any of the vpred models

@illyaeater any finetunes is completely serviceable. In fact, look at HassakuXL Illustrious which is based on ILXL 0.1, wasnt even noob tune, and its already looks so good. prompt adherence and image quality is complete different thing.

nah, that's a terrible model. There is a difference between civit "serviceable" and actually good to look at. When I say serviceable I mean no errors and looks at least the same if not better than using res closer to 1024

@illyaeater no error is prompt adherence. not image quality.

@illyaeater NOOB is a model trained based on a resolution of 1024. Forcibly using a resolution of 1536 will cause damage to the unet, severely affecting character proportions and composition. Even if you don't notice it, this is a fact.

@liangren26 you sound like you don't gen a lot of images, which is fair, but maybe you want to hold back on making definitive statements with no experience

@liangren26 I agree I notice the issues when forcing the size on something that isn't trained on it, it is very different than back then with doing older SDXL and hell even 1.5 I remember those days trying to test at larger sizes haha

What are the cfg and sampler you recommend?

It's a base model, it's not meant for genning with directly. Illustrious 0.1 looks pretty bad too by itself. This isn't an Illustrious issue, base Pony looks awful too as does NAI.

You really need finetunes, there are some for Illustrious 1.0 already and they look much better than the base model.

@Giangiotto your right I forgot ILXL 0.1 doesn't look so great on its own

@Giangiotto NAI looks amazing when you use its quality tags with artist tags

@scringle lol. No it doesn't. Every NAI gen absolutely reeks of AI slop.

@Giangiotto its better than illustrious v0.1 on its own, but your right it still needs to be fine tuned.

@Giangiotto you must be doing something wrong. you dont have to use (masterpiece, best quality, very awa) on every gen

Why spend money on your own site when you can use a formerly open source website and peddle your wares to a large existing userbase?

Not defending their approach, but isn't Nai4 non-local, paid-only? Also, Pony v7 is supposed to use the AuraFlow architecture as base, which, from what I remember, requires a lot more VRAM. A lot of SDLX users (like myself with only 8GB VRAM) simply won't be able to run an AuraFlow-based checkpoint.
EDIT: Correct me if anything I say is wrong.

It's not bad. It's not finetuned on a specific quality/style. So, it will generate average quality of whole dataset. Look at default illustious - it's the same, but after small (compared to the whole model training process) finetune it can produce very good results.

@Shio_N It's bad, I'm not the only one saying it, defending a company who makes money off stolen art assets should be a crime against common sense.

Is this also incompatible with 0.1 loras?

@roycolley861 Haven't tried, but knowing enemai, I wouldn't be surprised if they didn't work

ik i was looking forward to it man, apparently it has natural language prompting so we can use dandooru + natural English to specify what kind of image we want. I have been waiting ever since they released their paper.

@iKkYu_ReD Don't get too hyped about v2.0, since it's just gonna be a base model. The finetunes are what really bring out the potential in a model.

@ClownReplica that is true it won't really be useful until someone like WAI0731 gets their hand on it but either way i still wanna try it for myself

@iKkYu_ReD I really hope WAI fine-tunes one of the newer Illustrious models at some point, yeah. illustrious 0.1 was super basic at release and now look at how far they pushed that base model. Illustrious v1 is already a lot better than v0.1 at release, just imagine what the experienced fine-tuners could to with an illustrious v4.0!

@ClownReplica the v2 model is out on other platforms like tensorart and shakker but the model is not open to download, i think it would be better if WAI skips v1 and v1.1 and start working on v2 once's its out. I am no expert but i think it take a while and costs a good amount to train and release a checkpoint of WAI's caliber.

@iKkYu_ReD Newer illustrious versions releasing is the current biggest issue holding the model back.

Here: https://www.illustrious-xl.ai

This has a roadmap for planned illustrious releases. Why work on Illustrious v2.0 fine-tunes when v3.0 is to be released in a month? Right now, I'm guessing most fine-tuners and mergers are waiting for the most current Illustrious version to release, hence why there's been comparatively little in the way of v1.0 checkpoints.

@ClownReplica yes your probably right not to mention they are taking their sweet time to open the model to download as well. i guess i am going to stick with WAI's v0.1 for now.

@iKkYu_ReD The WAI v0.1 fine-tunes are the best available illustrious checkpoints right now, so yeah, same.

It's up on TensorArt.

@ClownReplica he did already a finetune if 1.0, it's called wai fcK ill, something like that,check his page

Honestly I would incorporate 1.1 into a merge model if download was available but I doubt it'll ever be tbh, so I suggest using the merges that are out their better or even NoobAI Epsilon which is practically a finetune of Illustrious 0.1 but isn't technically and it's way better since it was trained on more.

https://huggingface.co/OnomaAIResearch/Illustrious-XL-v1.1/tree/main

How long have they been gathering sponsorship money?

Also, why have a release roadmap (Illustrious XL | Next-Gen AI for Stunning Anime & Illustration Creation) if we can just buy the newest model for everyone?

@ClownReplica There is a reddit thread. Its about 2.1k bucks. There are already money in 3.5.

I think, they wanted money from every versions, but that didnt worked for the community. They should have released immediately the new versions, not just now. (Huge amount of loras for 0.1 dont work on later version)

@jerrywatkins159 Do you happen to have the link the reddit thread?

@ClownReplica https://old.reddit.com/r/StableDiffusion/comments/1jd8xq7/seems_like_onomaai_decided_to_open_their_most/

@jerrywatkins159 Thanks! They only opened the sponsoring less than 24 hours ago? Already being at 21%, there is a good chance we'll be getting our hands on v3.5 v-pred soon. Hopefully it lives up to what they are promising it can do!

@ClownReplica Illustrious v3.5 v-pred has already reached 100% stardust level and even higher. Their website literally says "We promise to open model weights instantly when reaching a certain stardust level (The stardust % can go above 100%)". If they were lying about this, it could be considered a scam because people literally paid for the model release. Hopefully, the model developers will keep their word and release the model weights soon. If the model is really as good as they say, I think people will be willing to buy their future models. https://www.illustrious-xl.ai/sponsor

@MindInTheDigits It was fast, some hours ago it was on 20%.

Now they must release it, or they will be pariah of AI community:)

@jerrywatkins159 @MindInTheDigits Yup, I just saw. Well, let's hope we can finally put this fiasco behind us, get an up-to-date base model with new bells, whistles, features and let the fine-tuners do what they've shown they are VERY adept at.

@jerrywatkins159 And it's now 105% Stardust, no release...

@Arctenox It's been like that for roughly 22 hours. The fine print implies that 100% is not actually the threshold upon which they will release the model. It's essentially a scam.
"We promise to open model weights instantly when reaching a certain stardust level (The stardust % can go above 100%). Different models require different Stardust thresholds, especially advanced ones. For 3.5vpred and future models, the goal will be increased to ensure sustainability"
It doesn't say that 100% is the threshold for unlocking. These scammers have some arbitrary goal set (or not, who knows at this point) at which they'll release the model.

Is it 2x what people have paid? 10x? 100x?

Who knows, give us more money!

@ClownReplica TL;DR. By the way...I was just stating facts of what it is at right now lol. I didn't need a whole explanation.

Also the whole thing saying it can exceed with no release is stupid in general.

Saying it can exceed but have it release at 100%, so you can still support the model makes more common sense in general. Instead of it exceeding past 100% with no release at all.

This is my opinion. Also if you didn't know V1.1 is available for download on tensor.art just you have to pay 10 USD, same goes for V1.0.

It not only says that but it literally tells people their funding goes to supporting open sourced releases, which as of today (22 days after your comment) is still categorically false. These people need to get their act together and do what they promised, or at least change their wording on the website and stop lying to people and falsely claiming their funds go towards open sourced software.
"Your Contribution Powers Open-Source AI for Artists & Researchers" - this is currently a blatant lie and misrepresentation on their website. IMO anyone who bought Stardust and donated to models that are not open sourced should at the very least be refunded.

lol. They should say v1.1 that's all they have) But that's what truth will give you instead of lies.

Seems like they're doing "take the money and run" tbh... rlly after someone pirated the V1.0 Illustrious and published it for everyone.

Tbh they seem to be not doing the whole opensource thing in general and they seem scammy after V0.1 Illustrious.

(This is all a personal opinion)

they are not training it yet

RealESRGAN x4plus denoise 0.5, steps 15

im confused... Last time NAI uploaded a new model was febuary the 3rd, wdym "frequent updates"?