I'm sure someone is going to come along and say "What did you expect for $500,000?" so I might as well slide in here now since I happen to be around. 😂

This resolution is way beyond the SDXL standard and it is quite common for it to not work properly. SDXL is still used entirely as a training base, and it was designed to maintain consistency at 1024px² or at resolutions very close to 1024px² (768 + 1280 = 1024px²).

Going a little over 1024px² won't be a problem, since this is the training of the entire SDXL 1.0 architecture, but 1536px² is over 50% of the model's base training, so it is completely expected that things will start to go weird.

For example, SDXL's base training makes it analyze 1024px² looking for items that are in the prompt. So if I type something like "1girl, solo, smile, 2 arms", it will look for these items within a radius of 1024px². Not that this is the limit of SDXL, but the base was trained this way. Now let's suppose that I increase the resolution by 150%, resulting in 1536px². When the image is constructed, the model will continue to look for items within a radius of 1024px² because it was trained this way. Let's suppose that the generated image shows a woman on the left side of the image, so when U-Net tries to cover the attention area, it will read the first 1024px², where there is everything that was requested in the prompt. However, when it continues execution, it will reach the right side of the image, where it will pay attention to just one more area of ​​1024px², leaving out 512px². Let's suppose that in these 512px² that are outside the attention block, there is the woman's left arm, then the model will understand that 1 arm is missing, and then it will add another arm to her, or it will simply reconstruct the woman a little further to the right so that the attention block can see one with two arms, but without changing the left side of the image where the 512px² that were left aside are. In the next loop, the same thing will happen, but it will only see one arm without a woman, so it will reconstruct her, and then when it goes to the other side, the same thing will happen, until the result is a completely deformed woman or two women with 2 arms.

The only way to solve this problem is to have a base training with high resolution, which happens in the most recent versions of SD 3.5, but is almost unfeasible for SDXL due to the amount of resources that the community has already invested.

Training Illustrious with more high-resolution images will not solve the problem, it will only reduce it a little. The solution would be to abandon SDXL and start training on the newer SD3.5 or Flux models, but the costs are too high for the community and it would take a long time to produce high-resolution images with the same accuracy that SDXL can produce at lower resolutions.

@AishaAI Everything you said it true and well articulated. I made my comment because the model's description says " V2.0 supports 512~1536 resolutions " , which (unless I misunderstood it) seems to suggest that 1536x1536 generation should be possible. In addition to this, the model can sometimes generate above 1024 Ok. I've done some test generations at 1596 x 1248 and they come out just fine.

@Shed_The_Skin I think they actually did some training at 1536px², but like I said, that would only help a little since the base SDXL was made entirely at 1024px²

The hash is C2A1A3EAA1 on both files, so they are the same

1 will blow up your computer and 1 will bless you with waifus

what if none are 🤔

It is better that you train a lora on the base Illustrious 0.1 model first. They say that further versions are lora compatible from those train on illustrious 0.1

这看起来并不像是一个易于解决的问题，例如“long hair污染long hair between eyes”的例子：我在Danbooru上检索了一下，在现有的帖子中有“long hair between eyes”属性并同时有着非“long hair”属性的单一角色数量（大部分帖子中存在多个角色分离两者）可能不多于10页，在现有的5.8k“long hair between eyes”tag的帖子中，实在有些少得可怜，作为训练数据集应该是不怎么够的。
个人认为在现在的Danbooru标签训练中应该没有训练让模型学习过Danbooru的tagwiki内容，换言之，AI大概是通过其自然语言处理能力来部分地理解tag词语的含义，然后将tag当作一种新的“语言”来进行理解，而非通过理解tag本身的含义来进行学习，要打个比方的话，ai对tag的理解大概相当于所谓的“散装英语”或者类似于日本的“偽中国語”？

如果能完全使用tag作为“语言”来训练模型/让模型学习tagwiki——抛弃AI的自然语言处理能力/进行优质的自然语言概念翻译，应当能解决这一问题——或者建立一个不会混淆的tag系统（类似于伪代码/元语言）用于训练，然后建立tag到tag的准确翻译——看起来无论哪种都需要不低的训练成本。

我就没见过不产生污染现象的模型，各种颜色，打个环境有XX鱼，贝壳，然后跑到人物头饰上。可能用#分隔会有点作用。

base looks bad, if you to check you can download from here https://www.seaart.ai/models/detail/cvbp2qle878c7387sppg

Ah, NOW they have their 2.0 able to be downloadable on civit, finally. On their official 2.0 page.. I don't know if it's the same as on seaart or another stable version. I'll have to see later

@Light7799 its the same, I have checked the hash.

They are both the same thing but a simple mistake while uploading during CIvitAi Maintenance, Blackout, Server issue and so on.... its not uncommon. 👍

@Kodokuna Ahhh, I see. Thanks for the explanation ❤️

This one is fine-tuned to improve stability. The other one is not fine-tuned for stability.

So, if you do LoRA training, you need to do it on the other one and run the LoRA on this one, as you will lose the stability fine-tune if you train it on this one. They have a blog about it on their website: https://www.illustrious-xl.ai/blog/7

I did some tests and this one is better, because it's literally more stable lol (visually)

Well, not just visually, actually

@Chris198 i've read their site and they do too have 2 versions of 2.0 there, this one is better at generating itself and according to them the other one is better at creating loras, making them more compatible with illustrious fine tunes.
Though i don't know for sure why you'd use illustrious 2.0 stable base instead of a fine tune

@YangDenn You're right, I was just trying this model because people were talking a lot about it and I wanted to know if it was because it had better composition or newer insights, but I didn't notice anything better than models like "WAI"

@Chris198 it has better resolution 1536x1536

Стилем. В стабильной версии есть относительно выраженный и повторяющийся стиль.

https://civitai.com/auctions

I think its more stable