其实如此

用V2的反义词，不要用通用

@Bloodsuga 大佬你早说呀，好多了，感谢

@Bloodsuga OK

@Bloodsuga 我的clip skip也是2，网上有说法是xformer，和xformer的版本，您有用--xformer和其他的arg吗，我还有用--no--half和--no--half--vae

@Bloodsuga 我测试了一下，还是一样，看来不是--xformer的问题，不同机器出不同的图的问题看来无解。

@Bloodsuga 再问一下大佬，您封面的图是原生就用1024x1536生成的吗？我尝试过无数次，即便不同的seed都是三头六臂甚至多人出现在一张图上。无法用超过768的分辨率生成正常的人物图，感谢大佬

eta (noise multiplier) for ancestral samplers 也会对结果造成影响

@Accepted733 这个不是原因，因为如果eta (noise multiplier) for ancestral samplers改了，png info里面会有显示，包括eta seed delta改了也会显示。

@Accepted733 而且我也知道为啥封面图片无法复现了，这个图信息里面有mask blur，证明这是一张inpaint过的图

之前版本的网络我按作者的参数弄，基本上是一样的。后面的模型没弄Clip skip\Mask blur等参数就是畸形的。

机子的显卡型号不同，显存不同，都会影响。。。。就算参数一样，也会变一点，不过，没试过同样硬件的PC，跑同一样参数，会是什么结果，有条件的同人，可以测试一下。

@Bloodsuga 请问您是什么显卡？

@Bloodsuga 我是台式机的3060ti，不知为何，全抄了prompt还是出不来你图的那种细腻和高清质感

這個留言真的學會了好多, 玩得很開心, 是時侯老一老了~

尺寸不一样也会影响渲染效果

我是用N卡跑的，完全一样的数据（利用PNG info再sent to t2i），结果还是和作者的图差别很大。不过我没有对图片放大，因为会爆显存。

@quinton_ho863 俺也一样

@ustakchan663 高清修复会很大程度影响这个模型的效果，用与不用几乎是两个风格了，可以试着放大比例调小，每次1张或张

@AIchemi3T 第五张预览图没有mask blur，但还是无法复刻orz

@Bloodsuga 怎么加你discord啊

这个网站有时候服务器压力很大, 会有一部分请求失败

@Bloodsuga @9ty9t 非常感谢

@Bloodsuga 可以教教我吗

lora这些，就是所需的数据和插件都是已经准备好的，我还专门试过有没有所需要的像6500lora这个tag的区别。

图片宽高比的问题，宽高比设置正确，太大也会乱出

说的数据直接用图片信息功能套的嘛，图片高宽比这些数据都是一样的。出的图差距实在很大。

我也是，模型参数，都一样，但是出来的效果，差别好大，，，，不懂是哪里出了问题

@Bloodsuga 我模仿的那张图，没使用lora

@Bloodsuga 这些都是一样的，因为是直接用图片信息功能，机器给套过去的，自己也反复检查多遍。只要是图片文件里的参数都是一样的。额外的lora，vae等等都是准备好了的。

应该高清修复的问题，你看的分辨率是高清后的分辨率

@z973 图片信息功能套的分辨率，和高清修复后放大分辨率，跟原图里数据肯定一样的，反复检查过没问题。

@2714425114241 我用的B站别人的整合包,没有问题

webui的版本不一样图库的数据也会不一样，作者的可能是老版本的

硬件配置不一样的关系吧

宽高一样吗，如果参数都一样，出的图应该是差不多的。出两个人可能是图太宽了。

@2994731975131 是完全一样的，添加了一个人的tag后还是会生成两个人，又试了下把宽度调低还是两个人

@2994731975131 又试了下，你说的对，我把宽度调的更低一些后终于变成一个人了，谢谢

最确切的解释，你去看png info，里面有mask blur 4，这个信息意味着这张图经过了inpaint，因此seed会变，并不是原始的图的seed，而且inpaint的分辨率跟着img2img走，是txt2img的两倍。因此你复制进txt2img后png info会错把inpaint的分辨率放进去，所以分辨率是不对的。由于inpaint后seed会变，因此你是无法复刻作者的图的。

SD2.0 is more powerful that's obvious, but it has been trained especially for safe content not for NSFW Images.

2.1 has a problem with way too skinny faces and other things

This is a model for stable diffusion. You combine it with Stable diffusion 1.5 and it can produce results like this.

@Bloodsuga thanks for the help, I'm still a bit lost and when I asked that question I still wasn't quite located in which program they were working

@tomberwick1984524 @tomberwick1984524 Thanks for the help, I will try to guide myself well to be able to make my own images and publish them

the VAE is embedded, meaning VAE and Checkpoint are in the same file

@Bloodsuga ty

negative prompt "big head, bad head proportion"

看你想用那个，一般用最新的。

@Bloodsuga Mega Sadge but understood

You can also click on an image, and the details are on the right side. That is the prompts and some of the specs. There is also a "Copy Generation" button in the bottom left of that screen. Copy and paste that into a text editor, and you can see a more complete list of specifications, which might include any settings that aren't at default, like Clip Skip setting and ENSD number. It will also show the model hash number for verifying you are using the exact same checkpoint. It may also indicate whether Contol Net was used, and the specs of Restore Face and Hi-res fix, if used. (In case you don't want to download the picture in question, & just want to try to get the same Lora's, embeddings, and hypernets, rather than duplicate the image.)

@Bloodsuga thx for the info :D

因为你的配置不够

因為要搭配LORA跟其他VAE，還有要多次做終極SD放大去增加畫質，不是單純用提詞而已

@Bloodsuga https://i.imgur.com/6WYpJva.png

@Bloodsuga  我不會發圖 不知為什麼要放URL 他總是限制了我

@yuduanji 跟配置没啥关系吧，应该是他vae或者lora

@pitterparkerri634 vae我不太明白是什麼，我已經下載了完美世界模型，Fashion Girl, Ulzzang-6500, 不知道還要下載什麼才可以做出那種質素的圖

@Bloodsuga 很感謝你的指教，我嘗試了你的設定，畫面確實有所提升，但對比你提供的白髮女子美艷程度還是差得遠，若要評級，你的是SSS級，我的只有B級，附上圖片https://upload.cc/i1/2023/03/29/2Eb8Tk.png

目前没遇到

@Bloodsuga i m not uisng hiires and cfg scale is default 7.

running into the same issue with a 3070 ti

set COMMANDLINE_ARGS= --autolaunch --deepdanbooru --xformers --force-enable-xformers --disable-safe-unpickle --disable-nan-check --api --opt-channelslast


this works for me.

@Docuei With you set: " set COMMANDLINE_ARGS= --autolaunch --deepdanbooru --xformers --force-enable-xformers --disable-safe-unpickle --disable-nan-check --api --opt-channelslast "
i have black pictures ;(
i9 -12900F + NoctuaRTX3070
;(

So this was an error on Civit that caused all models downloaded in a certain timeframe being borked. Just download them again now the problem's fixed.

They're doing a great job letting everyone know about it. /sarcasm

I had the same problem with Illuminati and other 2.1 models. I checked some forum and found a couple of possible solutions: (1) download the same model from Huggin Face and for reasons sometimes it works (2) go to Settings - Stable Diffusion and enable "Upcast cross attention layer to float32" (this worked for me). Hope this helps

I found that hires.fix fixed my problems

because the writer of the book 完美世界 likes girls with red eyes and white hair

以及https://civitai.com/images/439691?modelId=8281&postId=131736&id=8281&slug=perfect-world

本來，用同樣的模型、同樣的參數、同樣的提示詞，尤其是同樣的种子情況畫圖，應該會得到非常接近一樣的圖，但在不同版本的<Fashion Girl>影響下，色彩風格卻變得非常不一樣，甚至影響了整個構圖、人物姿勢。

面容是如何做到超仿真的？我做的都很卡通，就算用原图的提示词。

@callonger 要人脸修复：CodeFormer，及用“vae-ft-mse-840000”，如果用了動漫風格的Lora也會變得卡通化

I also feel like a misjudgment of the specified behavior. Like disgust, shyness

Need higher resolutions for faces overall. Are you using hires fix?

@sneakman thx, maybe I should try that

@Bloodsuga 原来如此，感谢解答

你说的“训练”实际上就是那个烧了八十万美金的Stable Diffusion初代模型，开源出来之后，后续对其进行fine-tune是一般民用级别的算力也能跑出来的

兄弟，学习AIGC是好事，但也要劳逸结合

@xujin_0413 笑死

what

https://civitai.com/images/731554?period=AllTime&periodMode=published&sort=Newest&view=categories&modelVersionId=65269&modelId=8281&postId=199392

I use Colab as well, and most of the free alternatives I tried were crap.

check out Yodayo

check out randomseed.co

This will happen if a prompt weight is greater than 2, like (arms up:2.1). Furthermore, there is a limit to the number of significant keywords that the model will consider, which I think is around 75. The more words, the less effect each word has past a certain point. There's a lot more technical explanation for this, but that's essentially what'shappening.

could you explain why?

What is clip skip?

How do I use that on EasyDiffusion?

i always see clip skip but in auto111 i dont see it as an option.. can you explain?

@iativic3 On auto1111, go to the Settings tab, then on the menu on the left select Stable Diffusion, go down the page, the last option is to change the Clip skip.
Make sure to APPLY SETTINGS on the top of the page!

What is the expected benefit? Thanks for info

thank you as a newbie i was scratching my head why i kept generating potato faces