Embeddings for SDXL
There are multiple embeddings here, pls check out each one of them.
Ditch those wordy prompts and get straight to the point with a single cue.
...and I see a lot of people are using SD15 embeddings on SDXL. Stop!!! They simply won鈥檛 work.
Reportedly SDXL no longer needs a long list of negative prompts, some models even perform worse when you use them. Below you can find the words for each embedding, select the one you want to use carefully.
I am going to expand this list over time. I've tested those prompts in each list, but there's always space to refine. And since some checkpoint are fragile against heavy prompting, I will try to keep these embeddings concise. Please check the list below and make sure they are indeed your desired effects before using them.
These embeddings are based on base SDXL 1.0 model, results may vary depends on what model you are using. But you can provide the prompts you'd like to use and request me to make embeddings for a specific model.
AC_Negs are general negative embeddings derived from negative prompts tested and recommended by AI Character in his article Here. You can read about more details. (Thanks man). Please check out his articles, they are super useful and informative. It is quite an extensive list, and I am not sure if you want to use all of them at once, so I split them into two parts for different usecases.
SK Styles are shared here by Sebastian Kamph (Huge thanks).
I have not yet put all of them as some styles are too niche, but it is definitely worth checking out maybe you may find something interesting for you. Negative prompts for each style are more or less the same, so I think it doesn't make sense to upload them individually here.
What are in these text embeddings:
ZipRealism = UHD, 8k, F2.8, RAW Photo, ultra detailed, sharp focus
ZipRealism_Neg = weird eyes, mutated hands, ugly, drawing, illustration, low quality, watermark, tiling, mutation, poorly drawn hands, poorly drawn face, bad anatomy, missing fingers, crooked teeth, extra fingers, disfigured, error, abstract, 3d render, cgi, fused fingers (sometimes when I type "deformed pupil, deformed iris" like I normally do in SD15, it often makes things worse...so they are not in the list)
Zip2D = masterpiece, best quality, ultra-detailed, sharp image, high resolution
Zip2D_Neg = worst quality, low quality, lowres, poorly drawn hands, poorly drawn face, bad hands, bad anatomy, 3d, blurry, watermark, signature, ugly, jpeg artifacts
AC_Neg1 = text watermark signature,low-quality,moir茅 pattern, downsampling, aliasing, blurry blur, glossy, jpeg artifacts, compression artifacts,poorly drawn,low-resolution bad ,error,pattern,beginner, grainy, glitch (More on poor quality image errors)
AC_Neg2 = distorted, distortion, twisted, excessive, exaggerated pose, exaggerated limbs, symmetrical,duplicate, pixelated, fake, hyper, overexposed, high-contrast, bad-contrast (More on style, composition and colors)
I do not recommend use all these negative prompt at the same time.
SK_3D = professional 3d model, octane render, highly detailed, volumetric, dramatic lighting
SK_Analogfilm = analog film photo {prompt}, faded film, desaturated, 35mm photo, grainy, vignette, vintage, Kodachrome, Lomography, stained, highly detailed, found footage
SK_Anime = anime artwork, anime style, key visual, vibrant, studio anime, highly detailed
SK_Cinematic = cinematic film still, shallow depth of field, vignette, highly detailed, high budget Hollywood film, bokeh, cinemascope, moody, epic, gorgeous, film grain
SK_Comic = comic, graphic illustration, comic art, graphic novel art, vibrant, highly detailed","photograph, deformed, glitch, noisy, realistic, stock photo
SK_DigitalArt = concept art, digital artwork, illustrative, painterly, matte painting, highly detailed
SK_Fantasy = ethereal fantasy concept art , magnificent, celestial, ethereal, painterly, epic, majestic, magical, fantasy art, cover art, dreamy
SK_CyberPunk = neonpunk style, cyberpunk, vaporwave, neon, vibes, vibrant, stunningly beautiful, crisp, detailed, sleek, ultramodern, magenta highlights, dark purple shadows, high contrast, cinematic, ultra detailed, intricate, professional
SK_Ghibli = Studio ghibli style, Art by Hayao Miyazaki, Anime Style, Manga Style, Hand drawn, cinematic, Sharp focus, humorous illustration, big depth of field, Masterpiece, concept art, trending on artstation, Vivid colors, Simplified style, trending on ArtStation, trending on CGSociety, Intricate, Vibrant colors, Soft Shading, Simplistic Features, Sharp Angles, Playful
SK_Vectorart = Vector art, Vivid colors, Clean lines, Sharp edges, Minimalist, Precise geometry, Simplistic, Smooth curves, Bold outlines, Crisp shapes, Flat colors, Illustration art piece, High contrast shadows, Technical illustration, Graphic design, Vector graphics, High contrast, Precision artwork, Linear compositions, Scalable artwork, Digital art (This one has a very strong effect, please use higher weight on your main prompt)
...more to add
If you have better prompts that you would like to share to the community or you want me to remake one of these embeddings, you can leave your comments down below.
All demo images were generated without adetailer and controlnet.
Like what I'm up to? Feel free to treat me a Ko-fi. Cheers!
Description
FAQ
Comments (44)
Thanks, kind of lacking embeddings for sdxl, https://github.com/Tropfchen/ComfyUI-Embedding_Picker
This helps if you use comfy
Any chance you might consider some embeddings targeted at kind of broad character based targets?
An example: I have a "Morphing" Lora that helps me replace parts of people in inpainting with whatever i type in "X morphing into Y". So i can make someones hand look like its made of fruit, for example,without losing the context and making the whole thing disconnect from the wrist. Its dodgy though, its more made for whole-person transformations.
One other Lora, that i don't have to trick is "Strange-head", but its only for sd 1.5. It allows me to replace someones head with objects, and it also tended to leak out to other parts of them if you set the prompts right ,which i liked to take advantage of.
Anyways, forgive me for the long explanation if its of no interest, but it seems to me any kind of model that could give you say "X combined with Y" (A man combined with a banana) or "X replaced by Y" or similar stuff.
I find that theres an emphasis on genres or styles on one end, and the other end is specific characters, or series, or artists. I'd love to see more tools designed to break the rules of SD. The models dont like to combine stuff together right now because a lot of the logic underneath seems based on keeping concepts separate in space, so that for example if you type "man holding strawberries" hes not holding them in his mouth, his hand isn't turning into a strawberry, etc. Essentially i think SD getting better at specificity is also making it a bit rigid. We need broadly assigned textual embeddings and Loras to try and counteract certain logic that would usually help.
Thanks for the suggestion, it does sound like an interesting idea. The only issue is that it will be a pain in the ass to collect right dataset for that. I will add it to the list, and take a deeper look when I run out of other interesting ideas to make. 馃挭
@NextMeal聽Thanks for the response, if i have any clever Ideas regarding a dataset i'll contact you! The dataset for the Morphing Lora I use
( https://civitai.com/models/140468/zeitgeist-morphingtransformation-concept-lora-for-sdxl )
seems to be based on the same type of data as the various "werecreature" Loras and Anthro models, i think its just tagged differently, and more ambiguously.
Thanks again and good luck with your work!
hi guys, i have a question, where do I have to put this embeddings, wich path? stable-diffusion/embeddings ? Because my sd doesnt detect them
it is stable-diffusion/stable-diffusion-webui/embeddings, but you also need to make sure you have the latest version of A1111
@NextMeal聽Oops! I thought I had the same problem. Went to download one of yours (first embed in weeks) and discovered my collection of embeds was gone. Downloaded yours and it's there. So, check that folder!
fixed, thanks. bad downloaded file
The latest version of webui also will filter on the model you have, but not if you change versions. If you were using SD1.5 and you change to a SDXL checkpoint, make sure you hit "refresh" when in the embeddings tab.
hey,
good collection ...
may you can take a look here ...
Thanks for the comment, your articles are indeed helpful. That's also why I kept the embeddings here short. as mentioned in the description here, SDXL doesn't need tons of negative prompts, but I still find them helpful if you are targeting a specific style, they can help you get straight to the point directly. The ones here are tested by other people or me and have been proven to be effective, although these are based on SDXL base model and I am not sure about their versatilities to adapt in each checkpoints. so far I find popular SDXL models on the site are very different. Unlike in SD1.5, default settings will almost always get you decent outputs. They all need different prompting styles, cfg, sampler, and steps for the best outcomes
there is an error: warning, embedding:ac_neg2,embedding:ac_neg1 does not exist, ignoring.
with the rest of the embeddings from the post there is no such error
Ok, I will take a look when I get time. I鈥檝e been using them normally for quite some time, so this is weird.
is an embedding = a word template?
in which case, writing the same set of prompts would be the same as just including an embedding?
I'm trying to understand, because if that's the case it seems it works like a dynamic prompt where you can call an external file that contains a prompt.
It depends, they can either be trained or mixed. In this model, these are mixed so they indeed work in the way as you said. Some popular SD15 embeddings were also created using this method, just not many people reveal their recipes. Another way to do it is to gather datasets or generate images for specific concepts, then train the embeddings to find what combination of tokens represent the concept best. pros is the file size can be further reduced and it can be used for training a non promotable subject, cons is the quality depends on the training
For realism at least, I find that I get significantly better results using only the negative embeddings (AC_Neg1, AC_Neg2, ZipRealism_Neg). Adding ZipRealism to the positive changes people to have a "plastic" look and feel.
Thanks for the comment. as SDXL is really sensitive to prompts, it is up to you to use what works best. 馃槉 generally I wouldn't recommend use a large pile of prompts all together.
"and I see a lot of people are using SD15 embeddings on SDXL. Stop!!! They simply won鈥檛 work."
I guess the reverse is the same thing? The SDXL emb are not retrocompatible with an SD15 model?
yep. there are two text encoders in SDXL and the vector size is not the same.
I'm having trouble understanding how to use these with mage.space. Mage.space now allows you to provide a URL to any model, embedding, LoRA etc and then use it.
I've imported many of yours (AC_Neg1, etc) but I'm not sure if I should be wrapping the keywords in <ti:...> for negatives or if something should go in the prompt to activate them.
Idk, I鈥檝e never used that neither馃様
@NextMeal聽Can you tell me if all of the triggers to into the Negative or do some of them go into the prompt? Once I know that I can experiment with if I need <ti:...> wrappers etc.
@TheZeke聽The ones with _neg postfix go to negative prompts, the rest go to positive prompts.
@NextMeal聽Thank you :-)
One suggestion. Adding "mannequin" to a negative greatly improves realism.
how do i use this in comfyUI? i downloaded the file to the embedings folder.. there is still one in for SD15 (easynegative).. what s next ?
(embedding:file_name:1) at the end of in the prompt
Now, just start typing embedding and it will offer you the list of those available.
They work very well for me by the way, thank you for your work.
Nice job.
Please, could you explain how you train a style ?
These embeddings were not trained, they are mixed. They are basically concated vectors.
The SK styles appear as embeddings so I put them in that folder, however when I put them in the negative prompt nothing happens. But if I put them in the positive if it catches the style
... That's because the styles are positive embeddings. While mostly used for negatives, TIs can also be positive, like in this case.
I am theoretically intersted in the "Anime" variant.
And the readme section for it mentions multiple styles. But... there's no "here's a sample" for each of them??
Please fill out the sample image part of the post with a sample for each available style?
what's the syntax to trigger different styles?
You have the embedding in the embeddings folder and put the name of the style in your positive prompt.
Look at the purple text in the info on the right hand side
Its funny how often we develop XL negatives and they're still using 1.5 embeds. XD
Somehow couldn't be used on pony anymore...
I saw several prompts with AC neg1 as not an embedding, but a Lora instead. Is it a mistake/misuse or there is indeed a lora,<lora:ac_neg1:1>?
where are any of the '-neg' versions of these embeds? TI or LoRa or otherwise...?
Thank you for your work, everything works very well in ComfyUI.
I copied/pasted your description into a note to help me find my way ;-)
ok, so where are the first few 'negs' versions? i looked at the example image for "AnalogFilm" which is in your positive and the AnalogFilm-neg is in the negative - there's only the positive embed to download though.... where's the neg?
Far from an expert but I'd just assume you use the positive prompt in the negative to eliminate the style from an image.
The cyberpunk model gave me an anime image and the anime model gave me a cyberpunk image. The General_Neg1 (AC) has worked for me in the most consistent manner so far.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.