Maybe in version 3.5 it will fix that issue. I Had the same issue when I typed in city and what came up was my character sitting on a desk hold and a tea cup.

Don't get me wrong. It's a good model and the way it does things and details it looks very great. But the fact that it ignores some important prompts such as color, background and other features. It makes you wonder if it has a mind of its own

Try to change individual prompts strength and see if it fix problem. example (1girl:1.1),(city:1.1)

Adjust number each time until it follow your prompt.

And don't use too much prompts, the more you use the less it follow them. Keep it simple and small for better results.

Black blotches? That is definitely the result of the lack of VAE.

bro I got this problem in the version2.5 too, I think its a common problem

Same problem, is there not a VAE included in v3?

I used this extension for the automatic1111 webui https://github.com/bbc-mc/sdweb-clip-changer

Use fp16; it's faster and uses less VRAM without noticible downsides.
fp16 (floating point 16 bits) uses less precise data structures than fp32, but unless you're training, merging, or trying to recreate something generated with fp32, there's no practical reason to use fp32.

Most open source community models are going to be f16. It's less precise, but faster.

Basically, if you're making anime waifus go with f16. If you're a scientist researching for your doctorate, use f32.

好像2.5更好，也可能是我不会用

请问3.0 vae是什么？

@galeony 使用自然语言写prompt，TI也更新了EasyNegative2

同问

测试多次原来还是 Counterfeit-V2.5.vae.pt才行！

you can use kl-f8-anime2

I'm sorry, but what is "normalized" prompt? Do you mean like human language prompt?😂

@NotAPersonJustACat to be fair, I don't know either lol. Hence, the double apostrophes.

You need the VAE. I like this one https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.ckpt

But I believe most use https://huggingface.co/andite/anything-v4.0/blob/main/anything-v4.0.vae.pt with anime (I like the detail of 840k).

Name the file CounterfeitV30_v30.vae.pt and put it alongside the main file or just enable the VAE selector in the webui. I like doing the later as it lets me swap them out (sometimes things like Deforum like specific VAEs).

@jonshipman with "alongside" you mean in the stable-diffusion models' folder?

@angy2218581 Yes

@jonshipman Thank you very much! I'll try it out. Do you know where can I read more about stable diffusion desktop? I would love to learn how to do better stuff

@jonshipman I downloaded the vae and placed it in the models folder with the same name as the original model, with the extension you said, but I still got different results, I am even using the same seed. I got this result:
https://imgur.com/a/oDhy3o2
which is pretty different :(

@luciodipre367 did you select the vae in the webui interface? You need to go to ->settings->stable diffusion->SD Vae drop down menu. Select the vae you downloaded and then click apply settings.

@DreamsOfElectric Hi, I'm really sorry to bother you but I'm new in this field and I really need someone's help to understand some important things. So, I have installed stable-diffusion v1.4 on my pc via Git and I run the program on the browser using the link. What I would like to know is does it change anything if I use a program like this above "Counterfeit-V3.0" which is based on 1.5 even though I have 1.4 ? I guess not, right? It just means it's based on 1.5 instead of 1.4 (can you clarify my doubt, pretty please?).

@angy2218581 hi! sry I'm afraid I don't know the answer to your question. My best guess is it won't be an issue but I can't say for sure. I don't think there's any harm in trying. What's the worst that can happen? The auto1111 webui might just not render images for you. That's it. Feel free to experiment but if you're having black output images or bad render results you'll know the reason why.

@DreamsOfElectric Thank you! I'll try to download the 1.5 then. Are you using that too?

@angy2218581 yes I am using 1.5. I have a friend who's way better at this stuff than I am and he recommended 1.5 over 2.0.

@DreamsOfElectric Thank you very much for helping me out :)

@angy2218581 The SD 1.4 you downloaded is just a model that your run through Auto1111, it isn't like an overall standard diffusion program or anything like that. This is also a model just like SD 1.4. The model's are all completely independant of one another and you can switch between as many as you like (if they are downloaded and placed in the folder). Basically, you don't want to actually be using any of the standard SD 1.4, 1.5, 2.0 etc models, but one of the Model's from here. Download a bunch of them that have different styles that suit want you want to make and try them out.

@Dunnas Oh, thank you so much for the clarifications!!! If I may, I have one more question. Regarding the models that are trained on (for example) Lora etc... how do they work? Do I have to download both models?

@jonshipman @DreamsOfElectric  I had the same problem, thanks for your help bro's!

aguante stable diffusion en español papa!!!

・I have utilized BLIP-2 as a part of the training process. Natural language prompts might be more effective.
・I prioritize the freedom of composition, which may result in a higher possibility of anatomical errors.
・The expressiveness has been improved by merging with negative values, but the user experience may differ from previous checkpoints.
・I have uploaded a new Negative Embedding, trained with Counterfeit-V3.0.
There's likely no clear superiority or inferiority between this and the previous embedding, so feel free to choose according to your preference.Note that I'm not specifically recommending the use of this embedding.

copied from huggingface. In general the composition and color palette are better, but anatomy is terrible, like 10/10 bad hand and distorted face.

@NotAPersonJustACat have you compared Counterfeit V2.5 and V2.0? If yes, could you tell which one is better in most situations?

I, too would like to know. There's no info on the huggingface page for it

https://huggingface.co/openai/clip-vit-large-patch14-336

Why the fuck does no one have this answer...?

Bro thinks im akinator

You can install https://github.com/bbc-mc/sdweb-clip-changer as an extension.

You might also need to modify it. I had to edit sdweb_clip_changer to change ".to(sd_model.cond_stage_model.transformer.device" to ".to('cuda')", so if you get an error about tensors being on the CPU, try that.

@Baughn thank you!!

@Baughn I'm inviting you to my child's first birthday

Figured it out: Use the link for the extension that Baughn gave https://github.com/bbc-mc/sdweb-clip-changer, once installed, reload the UI and scroll down to the settings. There you should see the CLIP Changer section where you should check "Enable CLIP Changer" and paste "openai/clip-vit-large-patch14-336" into the field asking to specify the CLIPTextModel. I applied it and the cmd window should start downloading some 1 gb+ model. While its downloading, maybe get the https://huggingface.co/hakurei/waifu-diffusion-v1-4/blob/main/vae/kl-f8-anime2.ckpt VAE. I got it from some other reply to a comment here and it seems to work well. Once the cmd window finishes downloading it will list all the things that are being loaded and "CLIPTextModel applied: openai/clip-vit-large-patch14-336" should be listed there. Enable the vae too by dragging it to the VAE folder and change it in the settings. I tried to replicate the images on display and they are basically the same with about the same level of detail.

@Schizolifting leave CLIPTokenizer blank? or paste it there too?

@smokeymillz alt account, yes that’s what I did anyway my bad

@Schizolifting @mohammadrajabwali Hey bud that's what I did, and it was successful, but now after one restart the Clip Changer section in settings tab is gone, you had a problem like that?

@StereoNostalgic218 No, I’m not sure what the problem is there. My bad bro I wish I knew how to fix that

@mohammadrajabwali No worries friend, it was a problem with Baughn's edit "cuda" suggestions, looks like reinstall and re-edit again fixed it. Thanks again for your explanation.

@StereoNostalgic218 That's why I said ''might", yes. It turns out to depend on whether or not you use --medvram / --lowvram, but as of a little while ago the extension handles it correctly, so nobody should need that hack anymore.

Draw Things不了解，Stable Diffusion (MAC也可安裝)靠的是NVIDIA顯卡算圖，基本建議是NV2060 4G RAM以上(4G算圖很慢，最好是30系列+12G)。另一種方式則是用GOOGLE COLAB / COLAB PRO 線上安裝算圖，但COLAB近期針對部分功能似乎開始收費了，安裝方式網路很多介紹

IPAD肯定不是很行，PC的高端显卡才够用，起码30系列

AI绘制图片最重要的一步是填写恰当的提示词，其次是配置恰当的参数。你需要参考模型作者上传的图片中附带的参数来填写和配置你的提示词与参数。

@yunleme 您好首先感谢你的回覆，但我有试过完全按照创作者的提示词及参数1:1的照着算上，但结果问题还是一样很简陋

@aaao0156357122 你应当发现了例图中有一个负面提示词是【EasyNegativeV2】，这也是一个被专门训练出来的嵌入式(embedding)模型。如果你不会使用这种类型的模型，那你应该移除掉这个提示词，并在负面提示词输入框中填写“(worst quality, low quality, normal quality:1.6)"这样的一段提示词。

kl-f8-anime2

没有写的话应该是直接整合进模型了（猜测）

@Le_malins thanks!

用 2.5的vae ,另外还看到 也有用 vae-ft-mse-840000-ema-pruned.ckpt

有成功的吗？评论里几个VAE都试过了，还是和原图不一样。没有原图的光感和细节。还是说我的电脑有问题？

要放於/models/Stable-diffusion/models/VAE

@Le_malins thanks

@summerlala123611试试 kl-f8-anime2

@summerlala123611 我也是，复现不出来

@seq2193 帅，

@Acode kl-f8-anime2和v2.5的vae都试过了，和原图质感都不太一样，2.5偏灰，kl-f8-anime2更重

试试这个vae：kl-f8-anime2

@wy89214356982 要放在 /models/Stable-diffusion/models/VAE

@wy89214356982 這個我試過結果變成模型那裏用不了

Use vae: kl-f8-anime2

Same

Same here

You have to use a custom vae.

- Download it from here: https://huggingface.co/LarryAIDraw/kl-f8-anime2/resolve/main/kl-f8-anime2.ckpt

- Move it into your models/vae folder

- In settings, go to stable-diffusion, and select it as your default VAE

- Enjoy.

@Hexamidine hey, how do you know to use this VAE?

@Rongorongo I just followed other people's instructions.

@53rdturtle I use vae but it reduces the color brightness

same

use runpod.io its not free but kinda cheap?

This model is trained way too hard in this matter, faces are really stubbornly shaded same way. I would recommend inpainting face with similar model.

@Ywinel Thank you for your response. I appreciate it.

It seems that I have been able to resolve this my issue recently by discovering Brightness Tweaker LoRA.

https://civitai.com/models/70034/brightness-tweaker-lora-lora?modelVersionId=74697

my sample is below

https://civitai.com/posts/231474

vae: gsdf/Counterfeit-V2.5 at main (huggingface.co)

其他都保持一致

用高饱和的vae效果更佳些，测试了一下

复刻不出来啊，光影和细节都会差一些（问题找到了，是EasyNegative不对，要用V2）

kl-f8-anime2

@epsychic V2版本在哪里呀 找不到能发个链接嘛

@linyingaiacgn57782 gsdf/Counterfeit-V3.0 at main (huggingface.co)

me too

应该是环境的问题，我刚刚重新部署了一下，这个问题就好了。。

刚刚只是错觉，，还是有，无语了

我也有……蜜汁荧光绿块

我也是，找到解决办法踢踢我

我的也有，而且圖片感覺被 deep fried 了

也遇到这个问题，有知道怎么解决的告诉我一声

是不是用了badhand embedding

同遇到这个问题，有没有解决方法

The license for any given model is always linked below its reviews on the right side of the page:

https://huggingface.co/spaces/CompVis/stable-diffusion-license

11. Accepting Warranty or Additional Liability. While redistributing the

Model, Derivatives of the Model and the Complementary Material thereof,

You may choose to offer, and charge a fee for, acceptance of support,

warranty, indemnity, or other liability obligations and/or rights

consistent with this License. However, in accepting such obligations, You

may act only on Your own behalf and on Your sole responsibility, not on

behalf of any other Contributor, and only if You agree to indemnify,

defend, and hold each Contributor harmless for any liability incurred by,

or claims asserted against, such Contributor by reason of your accepting

any such warranty or additional liability.

That said, nobody likes a grifter my dude, I recommend keeping free stuff free

@UnsignedLongshanks I think the OP was talking about the images produced with the model.

配合一个VAE就好了

Try a new model or the normal model, emanoly pruned???

You have to use a custom VAE. kl-f8-anime works well.

@Hexamidine That worked really well! Thanks! <3

If anyone has trouble installing VAE, then use this https://civitai.com/models/23906/kl-f8-anime2-vae with this https://stable-diffusion-art.com/how-to-use-vae/

@Hexamidine Will reduce the saturation of the image

@VanOK Will reduce the saturation of the image

ı have this problem too how to u fix it?

me 2

same

orangemix VAE를 사용해보십시오. 원본 vae를 사용하려면 고해상도 업스케일을 사용하십시오.

更換別的VAE試

Same here. Seems like the issue is linked to the VAE embedded in the model. Just like another comment above, manually switching to other VAE resolved the issue for me although it might possibly affect the artstyle of the outcomes.

画布太大ai不会画整体了。用分块

分辨率太高是会出现多人多手多脚的情况，这跟AI训练的一些东西有关，所以还是先绘制低分辨率的，然后在调高比较好

可以尝试高分辨率修复，或者图生图来提高分辨率

别上来就大分辨率，用高清修复会好点，512*768的图基本不会有结构问题，再往上直出大图，增加高度就会出现人体结构错乱，增加宽度就会人数增加

先做小图，然后高清放大，你直接做太大的图AI会觉得你要画两张，于是就画了多个人

也有可能是你提示词太少了，AI用现有的提示词填充完画布还有多余的自由度，就会开始自由发挥。

可以尝试：1、增加提示词数量，限制AI发挥空间；2、减少画布大小，如果想要高清大图可以后面在图生图放大

分辨率问题，训练素材的分辨率一般都是512*512or512*768的，你可以使用高清重绘来放大分辨率

这种情况是分辨率太高了。一般画图长边不要超过512，想要高清图的话应该用高清修复功能，而不是调高分辨率

感谢兄弟们的解答。谢谢谢谢

分辨率太大了，追求高清可以看看这个视频，【拒绝低质量出图！3种快速提高AI绘画分辨率的方式，十分钟讲完！| StableDiffusion WebUI 保姆级攻略·高清修复优化细节无损放大教程】 https://www.bilibili.com/video/BV11m4y12727/?share_source=copy_web&vd_source=cb4a6bab509434517338c600fc983ebd

画布越大你的prompt就要更精确。如果人数大大超过你的预期，你可以尝试用一些除了人之外的东西占据空间。

负面词: extra head, extra body,

试试反向加权重，比如 Negative prompt: (Umbrella: 1.4)

@53rdturtle nice 解决了 

did you ever figure this out?, found it on hugging face, no idea what it has to do with this checkpoint though.

Here is how https://civitai.com/models/4468/counterfeit-v30?commentId=117932&modal=commentThread

@6framejab That thread no longer seems to exist. Also, does anyone know if CLIPChanger works properly with 1.6.x and the Stability Matrix launcher?

Pics?

generate the image u want, then whatever image u like, make sure you're using the same seed and regenerate it but this time using hires.fix and using the anime upscaler. Change the "upscale by" value to 2(multiplies ur current resolution by 2), do a .5 denoise and 0-10 steps

840000会好点

https://s2.loli.net/2023/06/16/y8RKXuFmavbEYJQ.jpg

@higekibaka 非常感谢，实测一致 OVO！

有没发现直接用例图的 提示词 生出来的图会比例图少非常多的细节，花朵，衣服装饰物之类的，画面变的非常平淡，试了好几个例图的提示词都是这样。

@CHINGEL 可能没有启用hires.fix（高清修复）？，推荐高清修复步数15，重绘幅度0.57~0.6，算法选Latent系列

@1097611730304 我复制的例图的提示词，高清修复基本上我出图就开的，修复步数、重绘幅度这些例图提示词里都包含了的，就很费解，用lora:add detail 会好一点，但是不太协调

@1097611730304 开了高清修复之后出图好慢，是不是显卡不行，还是设置的问题

https://huggingface.co/gsdf/Counterfeit-V3.0/tree/main/embedding

@Philuu thank you

trust me v1 better, v2 just add fix hand and may change pose if you used hand-neg already don't use it

我也是太鲜艳，效果很差，为什么

看着需要外挂VAE

可能是用的插件问题，也和主体画风有关系

我倒觉得这样可以主次更清晰

我也刚了解，应该是一个嵌入式模型，快捷写提示词用的吧

如果我的理解没有问题的话，EasyNegative是一个提示词合集，下载改嵌入式模型后丢到根目录的embeddings文件夹，之后直接在负面提示词里写改嵌入式模型的文件名就好了

https://civitai.com/models/7808

这是一个Embeddings模型，下载放到embeddings文件夹就好，地址：https://civitai.com/models/7808

C站都有的 负面词的embeddings

非常感谢各位大佬

It fixes that issue, but it gives me results with dull colors, and before, even though it had green spots, the colors looked vibrant

did you fix it? same thing is happening with me.

try something in VAE, it's have some error when you set it defaut by the way

thanks, but it's so weird that such a famous model have this issue , while other common model don't need specific vae. :p

Hey, thanks for the tip.
I need to change these lines in the /extensions/sdweb-clip-changer/sdweb_clip_changer.py?

A late response, you've probably already figured it out. It's 'down jacket'. To be more specifc like in the image you sent it would be '(your initial prompt) open down jacket, white sweater (the rest of the prompt)'. Winter tags do help. Do not specify it too much though (sweater underneath down jacket / down jacket on top), it'll make things worse.

yes, maybe u should choose another VAE model like 840000 other than animevae if u want pics more colorful.

有的没vae是会这样的

推荐一个VAE，Counterfeit-V2.5.vae.pt

did you find the answer? it's half the size of the normal fp16

@HongYue it looks like a pruned version, almost identical results

我在留言中没找到啊，想问下怎么用CLIP changer，谢啦

Are you on A1111 1.6.x? Were you able to get the Clip Changer extension to successfully download the new clip? I use Stability Matrix as my laucher/installer as they discontinued the old 1-click installer, and the console did not show it downloading the replacement CLIP model. Can you perhaps provide any assistance? I don't really want to have to go back to A1111 v1.5.x if at all possible.

Also, does the clip changer even support Xformers? that might be my problem.

Possible issues: restore face is enabled, image burning. With latter you can try antiburn extension

100% is the vae, you need to change your vae on settings, the one on the left sure is anything v4 vae.

most common used vae is, vae-ft-mse-840000-ema-pruned.vae

The VAE issure, i changed it to the vae used in counterfeit-v2.5, it looks more similar to the original img

If someone try to find how to make it just work - i finded video guide about it -
https://www.youtube.com/watch?v=Dw4LiGgR3hs&ab_channel=StableDiffusionenespa%C3%B1ol

And make sure you set vae from automatic (comment above) to vae-ft-mse-840000-ema-pruned.vae (AND APPLY SETTINGS)

The CFG scale probably needs to be reduced to 3.5 - 5.5

Hello, have you resolved it? I have the same problem as you.

@fangmGUge use vae, and don't use 2.0 value on negatives "bad quality" and similar, usually max stable is 1.5, anything over that can cause inestability

there is an icon when you create the image that is at the right upper part of the image there you can dowmload the image

if you generated it on your pc you already have it downloaded in
\stable-diffusion-webui\outputs\txt2img-images

that is troubling, where did you hear that?

nvm, i guess it was for midjourney only, or some program like that.

While I don't fully know the answer myself, but in Automatic1111's web-ui, under the "Interrogator" tab), the CLIP Model dropdown selector has a value of "ViT-L-14-336/openai". Running the interrogator with that selected produces a semi-conversational prompt instead of a booru prompt, so I'm guessing it has something to do with the kinds of prompts the author recommends to use with this.

+1

+1

change vae

+1

change your vae, klF8Anime2VAE or vaeFtMse840000EmaPruned will solve the green strokes

+1

Now I will download and try it out, I hope you are right, because so far all of the above I can only say about Pony. Really smart model that understands points of view, poses, perfectly understands the weight of different tags. She has only one drawback that there are not enough LORA for this model. For example, there are almost no Honkai Star Rail characters. The other models that I have so far tried, no offense to the authors, the most stupid and random, to make them draw something correctly is very difficult. I enter the correct tags positive, negative, put easynegative, and it draws normally, but only something simple. You want to make it a little more complicated? Like increase the weight of a certain tag? Get twisted bodies, hairy nipples, several bodies, strange elements of clothing, or very strange landscape elements....

I apologize for the large wall of text. I just wanted a place to vent))

@Sanko_96 every tag is unique so adding weight might work differently but generally it's in the 0.8-1.3 range. Below or above might result in heavy distortions. As for more complex things - it's all about ControlNET + Inpaint + Latent Couple, unfortunatelly. SD still can't do really complex things on its own without those tools.

me too

Maybe you should download the 9GB file, which can be found by my M1 Max Macbook.

maybe a sampler issue, do try automatic1111 since it's a bit less convoluted and less chance of user error. It can also copy the prompt with "png info" and then you can see what you might be doing wrong. This model is very good, even in the age of AnimagineXL and PonyXL I keep using it.

The images made using the onsite generator look pretty bad. Is there something we need to add like a vae or something?

Eyes issues often comes from low quality VAE decoder. Use official VAE checkpoint which if full size 327Mb
vae-ft-mse-840000-ema-pruned
vae-ft-ema-560000-ema-pruned

(maybe will be useful for others)

following

解决了吗

同样问题，做不到卖家秀那种颜色

可能是VAE的问题，尝试切换成kl-f8-anime2试试 https://civitai.com/models/23906/kl-f8-anime2-vae @hardforaname146 

@Nen1y 捕捉大佬！！！最近一直在看您的课，收获颇深，帮助我解决了做独立游戏缺少美术素材的难题 qwq

@Nen1y 竟然见到了大佬本佬，赛博磕一个

你可以试试用ai 帮你生成关键词

I've managed to get some decent results out of these settings on ComfyUI.

Steps: 35, CFG: 5, Sampler: dpmpp_2m Scheduler: Karras

basic prompts seem to work fine with it. check out the images under the bot and look for prompts to work off if youre unsure.

Hope this helps

Use another VAE (e.g.: 840000 ema) to fix it. They should update the description with this information asap

dont care

i get where you're coming from but just imagine the amount of images you gotta crop one by one, it takes a long time, and its unecessary work when you can just remove them with negative prompts

just using negative prompt to avoid this, nude

Cropping takes no time at all, even across 100s of pics if you know how to use editing software properly (then again, I guess they wouldnt be AI users if they did, hur dur). Also, the opinions of those who rely solely on negative prompts to fix all their problems should be completely disregarded lmfao

Happy New Year, Mate
Hope you are doing well :)
I appreciate that you point out this, but I have trained loras and some models but it's really exhausting thing to crop, adding tags, choosing correct resolution for image out of 600+ images removing watermark in every image is not possible as some image have crucial details there and i can't be removed.

Hasta tu propio modelo prr@, todavía que lo hace gratis, andas gritando como prr@