介绍(Chinese Version)
简介
此文本嵌入(text embedding)为 负面文本嵌入。它能够在对画风影响较小的前提下改善AI生成图片的手部细节。如果它让你的模型表现得比以前更糟,请勿使用它。您可与其他负面文本嵌入一同使用。虽然它是为 AnimeIllustDiffusion 模型设计的,但您也可以在其他模型上使用。
它似乎在较高的 CFG Scale 下(>=11)表现得更好。
如何使用
对于 AUTOMATIC1111 WebUI 用户
下载模型文件 badhandv4.pt;
将模型文件放置于您 Stable Diffusion WebUI 目录下的 "embeddings" 文件夹内;
打开 WebUI,在负面提示词框内填入 “badhandv4” 以触发效果。
对于 ComfuUI 用户
下载模型文件 badhandv4.pt;
将模型文件放置于您 ComfyUI 目录下的 "models/embeddings" 文件夹内;
打开 ComfyUI,在 负面提示词框内 填入 “embedding:badhandv4” 以触发效果。
SDXL 版本
我将 badhandv4 移植到了 SDXL 格式的 embedding 上。下载地址参见:
Negative Embeddings - AIDXL Series Models - v0.4 | Stable Diffusion Embedding | Civitai
Introduction (英文版本)
Introduction
This text embedding is a negative text embedding. It can improve the hand details of AI-generated images with less impact on the style of painting. If it makes your model behave worse than before, please do not use it. You can use it with other negative text embeddings.
Although it was designed for AnimeIllustDiffusion model, you can use it on other models as well.
It performs better with higher CFG scale (>=11).
Quick Start
For AUTOMATIC1111 WebUI User
Download model file badhandv4.pt;
Move the model file to the "embeddings" folder in your Stable Diffusion WebUI directory;
Lanuch WebUI. Enter "badhandv4" into the negative prompt box to trigger its effect.
For ComfyUI User
Download model file badhandv4.pt;
Move the model file to the "models/embeddings" folder in your ComfyUI directory;
Lanuch ComfyUI. Enter "embedding:badhandv4" into the negative prompt box to trigger its effect.
SDXL version
I converted badhandv4 to SDXL format embedding. For downloading, see:
Negative Embeddings - AIDXL Series Models - v0.4 | Stable Diffusion Embedding | Civitai
Description
FAQ
Comments (65)
We need a safetensors version of this.
Safetensors are un-supported in the webui, its either .pt or .bin currently
@eye_am_bored You are wrong on that, I am literally using a .safetensor embedding (EasyNegative) for over a month now.
This is a 18.9Kb file... what's the point optimizing something that small ?
@tetsuoo The reason is that, unlike a safetensors file, a pt file can contain malicious code that cannot be fully checked automatically.
我还以为他会给easynagetive一拳呢,没想到相处的还不错
In which folder should I add this file?
The 'embeddings' folder in your stable diffusion installation directory.
@EG Thank you very much! ♥
This is a really good embedding, but could you also make a safetensor variant?
You cant get pickled by an embed because there's no pickling lmao
@loktar Can you explain why in more detail?
Looking at all the images people are posting using this embedding, they tend to have messed up hands, so does this actually do anything? Writing extra fingers, bad hands, and so on forth in the negative prompt doesn't really improve hands, so how is this improving them? From what I see, the hands still look wonky in 99% of the images posted.
You posted a pertinent question. The answer is - I can't guarantee that it will always work, because it is almost a black box.
Here are my observations:
It was trained on the AnimeIllustDiffusion model with about a few hundred images of extremely malformed hands. The learning rate is high and the training converges. Therefore it does not overly affect images outside the training set, which means that if the images have mildly distorted hands then it will not help much.
Unfortunately, there is no effective solution available now. This is tradeoff. If you want to target fix the concept of 'hand', you need to do a lot of supervised learning with a very low success rate.
I have produced dozens of text embeds for restoring hands. Those that have been effective in restoring the hand have invariably changed the original image - to the point where I think they have coincidentally restored the hand by replacing the original image in its entirety. badhandv4 is the only one that has managed to slightly correct the detail in the hand without damaging the original image.
If you really want to solve this problem you have to do it manually with controlnet
I doubt this one really works.
我也怀疑
It actually does.
I came here because I tested the few bad hands embeddings I had and this one won the comparison hands down. So, I came here to see if he's got more embeddings. He does. No, I didn't test them yet.
That said, not every embedding works well with every model or in combination with every other embedding. Always test what combo is best for the model you're using.
Eh most of these are placebo
They are gonna mess with the composition either way and might get you like 20% more gens with ok hands but it wont magically give you good hands every time or even every other time
@r6zhtfg that’s mainly due to the fact that AI-generated software doesn’t fully understand what the word “hand/hands” mean. Hands are super complex even for ai to render, so until that’s fixed, generating near perfect hands every outcome isn’t rlly possible, but at least w a LoRA like this it’s more likely.
没用呢老铁,生成图片信息里显示没有调用这个
i keep seeing 'anime_badhandv4' and don't know if that is a different file than this one.. i've searched all over for 'anime_badhandv4' and this is the only thing that comes up.
I guess no one else named his model "badhandv4". So it's the same.
anyone know bad hand v3?
If it refers to earlier version of badhandv4. No such thing :D. But this may help: Useful Quality Embeddings - AnimeIllustDiffusion - aid291 | Stable Diffusion Textual Inversion | Civitai
@EG thx!
unlike negative_hand, this badhand try to "hide" hand, "not fix" and may change your image
please add sdxl version
Yes. You may refer to: Negative Embeddings - AIDXL Series Models - v0.4 | Stable Diffusion Embedding | Civitai
in negative: (((internal-organs-outside-the-body, internal-organs-visible, anatomy-description, unprompted-nsfw ,worst-human-external-anatomy, worst-human-hands-anatomy, worst-human-fingers-anatomy, worst-detailed-eyes, worst-detailed-fingers, worst-human-feet-anatomy, worst-human-toes-anatomy, worst-detailed-feet, worst-detailed-toes, camera, smartphone, worst-facial-details, ugly-detailed-fingers, ugly-detailed-toes,fingers-in-worst-possible-shape, worst-detailed-eyes, undetailed-eyes, undetailed-fingers, undetailed-toes)))
in positive: masterpiece-anatomy-perfect
never even once failed me in generating perfect hands and anatomy in general....
didn't work for me, still got mutated hands with 20 fingers. Ruined an image that otherwise looked good.
doesn't work for me. probably depends on what base model you're using.
@fablegenius go on my proficle and test my articles. those are shorter and work way better. this "collection was from my early days. won't delete it, because it still is useable for some people
Looking through your image history I see alot of bad hands...
@Enlinze lol!!! yes, because not every image had negative prompts on them. and, like you said, it's "history". i published already a lot of better prompts on my articles. you should take a look at them.
i won't delete my last comment here, but there is a very easy prompt to get everything correct. all you need is this in negative
inequalities:2
Don't you recommend using: "solid-best-physical-appearance-perfect::2"?
every time you write a comment you have something "better", you just keep spiting up random words
@elguachiiii yes, because i don't give up on testing new prompts. no one else here on civit does this. all they do is creating those shitty embeddings with a bunch of prompts that hopefully can generate something good, but not a single prompts that can do all of this together for the whole body.
@stapfschuh
I personally don't support the use of lengthy text prompts as a template.
The CLIPs of different models are different. Therefore, the same prompts are not universal on different models, especially those very long prompts. Maybe you happen to write a prompt that's perfect for one model, but it doesn't usually have the same effect when you use it on other models.
The text prompts you enter can be recognized by the model if and only if the concept is annotated in the training data of the CLIP model, AND when CLIP explains it to the diffusion model, the diffusion model can also understand it. Otherwise, no matter how long the prompts are, they are just random tokens.
Embedding models and prompts are essentially the same. In principle, all input text prompts will be encoded as embedding and sent to the model. Therefore, embedding is actually the packaging of the text prompts. Any prompts can be packaged as an embedding - if you want to share your negative prompts with others, you can also directly package it as an embedding file for others to use.
In addition, the effect of text prompts is limited. The modulus of the token vector in text prompts is approximately 0.05, and the manually produced embedding model can reach any value. Even if you use a extremely long prompts to control the model, the effect is still weak in comparison. For those bad models that usually generates bad hands/anatomy, text prompts can only "weaken" bad things but can't "eliminate", or even strongly reduce them.
I don't think the embedding model can solve all problems, but I can't agree that long prompts can do a good job.
@stapfschuh i never use negative embedings an usually keep negative promt short and just add specific negative if image is deformed, no universal ones
Interesting. Will have to test this.
I was actually kind of surprised this works as well as it does.
@sansmia visit my profile and read those articles i published. there are more ways/prompts for better images.
@Euge_ "i wont delete my last comment, everyone makes shitty embeddings, i leave up even the ones i want to delete for others" - signed, (deleted)
@Legend_of_Chavo_G after telling us to visit his page
太棒了,兄弟!
非常的哟西
兄弟你真牛,我在韩国上课教授也必须用你的《坏手》
Works almost like magic, so far has given good hand results in most checkpoints I use :D
Wow!
有人可以解释下embedding的作用原理吗,为什么模型加了这个后,生成的手的效果就会变好,embedding不是提示词的集合吗?应该不需要训练得到吧?生成的效果不是应该取决于模型吗?
embedding相当于把一段提示词打包起来,用向量代表这段文字。但是,文字的表示效果有限。所以,把这个embedding向量保存然后用图片训练来调整向量数值,这个embedding向量就更能描述训练的图片。
比如坏手,模型能生成坏手但我们却不能用一段文字准确地描述什么是坏手,就需要用坏手图片去训练一个embedding模型来表示它。得到的embedding模型代表了一段描述坏手的提示词,把它加进负面提示词后阻止模型生成坏手的概念。
非常不错的嵌入
请问这个版本在pony上能用吗?
Why is this making all my characters blue-eyed?
uhmm,do you still have the download link of the sdxl version?the link in the introduction is missing now :(
为什么我下载之后不行,没有在列表里面看到,是不是少了一个东西
comfyui里用了cfg11,经常还是出现坏手,头痛
ber,用于SDXL的怎么无了
Details
Files
badhandv4.pt
Mirrors
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
16993_20068_badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
6.badhandv4_[Mar 08, 2023].pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4(1).pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4-AnimeIllustDiffusion_badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
learned_embeds.bin
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4-AnimeIllustDiffusion_badhandv4.pt
badhandv4-AnimeIllustDiffusion_badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
bad_hand_v4.pt
badhandv4AnimeIllustDiffusion.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
badhandv4.pt
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.

