Negative Embedding for Realistic Vision v2.0

Negative Embedding for Realistic Vision v2.0 - v1.0

NSFW

This is a 75 token negative embedding for the model Realistic Vision v2.0. It should you help attain a more realistic picture when prompting. All training images were generated from that model, and all the training was also done on that model. The embedding will likely work on other models, but I have not tested it on other models.

To use this embedding you need to download it, put it in your embeddings folder, and then activate it in the negative prompt by typing out it's filename in the negative prompt. By default that will be "realisticvision-negative-embedding" (do not enter the file extension in the negative prompt). If you don't like that activation phrase, then you can change the filename to whatever you want, and that becomes the new activation phrase.

Description

Trained for 5342 steps on 5230 512x768 images (I hit stop a bit too late heh). The basic training parameters are 0.005 base learning rate, batch size 1, 1 grad, CosingAnnealingWarmupRestarts Scheduler activated with a 500 step cycle, advanced adamW parameter options were activated but not otherwise touched. No captions were made or used, style.txt was used as the prompt template file. I ran 1 batch because I decided to turn off xformers, and all my card could hand was a batch size of 1 512x768. Barely. The extension I used for training was this one.

Due to the content of the training images I will not be uploading them, instead I will point to the reddit post from which I derived the training image prompt: link. In order to generate the training images you would need to put the negative prompt into the positive, put the positive into the negative, and remove all subject focused words. All other parameters should be the same- such as the step count and cfg, sampler and so on.

FAQ

Comments (21)

Rin_chanApr 10, 2023· 9 reactions

CivitAI

It's nice to have negative embedding for realistic models... I don't get the examples tho... I mean, both guys are black and white, and bunny looks not better or worse. I would honestly prefer to see hands or legs or disconnected architecture fixed.

SeekerOfTheThicc

Author

Apr 11, 2023

the examples are the same seed/settings/prompts as the four preview images for realistic vision v2. I chose to do a comparison to those so people would 1. know that I wasn't cherry picking results, and 2. make an informed decision about whether they want to download this negative embedding

dpfdpfApr 11, 2023· 13 reactions

CivitAI

Can you explain a bit more about what this is intended to do? The examples look more like different seeds than "better" or "worse". Is this supposed to be an add-on for RV, or a replacement, or?

ash0080Apr 11, 2023· 5 reactions

CivitAI

Haha, could it be that there's something wrong with my eyes? I think there's been a significant improvement, guys. Take a closer look at the rabbit's eye details and the facial features in the pictures - they're so much clearer! I must say, the author has done a pretty badass job with this!

ezez_Apr 23, 2023· 4 reactions

CivitAI

For me, using it makes the results way worse ... I'm using RealisticVision2.0 (e6415c4892), do you know what wrong do I do ?

secnepJul 6, 2023

I'm finding it accentuating muscles on bodies, personally. Although, the author recommends using it with their model, which isn't what either of us are using.

gdav6owj10Apr 23, 2023· 4 reactions

CivitAI

Very cool, but when I use this, other negative prompts get ignored which makes it kinda useless.

When I prompt "apple" and negative prompt "red, realisticvision-negative-embedding" i still get red apples all the time.

Is it because the embedding uses all 75 tokens? But i thought A1111 is able to process more...

openfireMay 7, 2023

that's interesting did you test this on other negatives embeddings as well ?
did you tried what happens on different samplers ?

SeekerOfTheThicc

Author

May 8, 2023· 4 reactions

I didn't see this comment until now, sorry. Webui can process many tokens, but yes it will often overpower smaller token words. I look at it sort of how I look at voting- the embedding is 75 tokens, versus the 1 token for red. 75 is much greater than 1, so if there's is a conflict, the 75 would win. Under the hood for webui all I know is that starting back in... December or January, webui sends tokens in chunks of 75 (actually 77 I think because there is also a token for start and end), and there is some sort of averaging occuring.

To combat this you can weaken the embedding magnitude (realisticvision-negative-embedding:0.8) for example, and/or increase the magnitude of the other words in the negative prompt- (red:1.4). I find lowering the power of the negative embedding below .8 or .7 produces... weird results that are unhelpful, and I rarely have gotten a useful result my increasing the magnitude past 1.4 for normal tokens.

liftweightsMay 12, 2023· 1 reaction

@SeekerOfTheThicc I have found that using your embedding with a weight of 0.3 and a CFG scale of ~3 can produce some fascinating results with inpainting, especially in conjunction with ControlNet.

dreklyMay 9, 2023

CivitAI

@SeekerOfTheThicc This is incredible... but in a strange way.

When I use your negative prompt, the Lora I trained looks exactly like the person I want.

However, when I remove your embedding, it turns into a blurry mess and looks nothing like them?!

How does it have such a huge effect on the Lora I made?

SeekerOfTheThicc

Author

May 9, 2023· 2 reactions

The methodology I use for training a negative embedding essentially takes a negative prompt and magnifies and focuses its power beyond what you can get from just typing up a negative prompt.

It sounds like you may need to re-examine the workflow and settings you used to train your LoRA. I'm not experienced enough in training LoRA's to be able to give better advice than that, but I wish you luck. It takes patience and plenty of repetition.

dreklyMay 10, 2023

@SeekerOfTheThicc would it be possible to tell me the negative prompts you used so I could experiment and see which one makes my lora work correctly? lmao

SeekerOfTheThicc

Author

May 11, 2023· 4 reactions

@drekly Yeah. The negative prompt involved was:
asian, naked, nude, out of frame, tattoo, b&w, sepia, (blurry, un-sharp, fuzzy, un-detailed skin:1.4), (twins:1.4), (geminis:1.4), (wrong eyeballs:1.1), (cloned face:1.1), (perfect skin:1.2), (mutated hands and fingers:1.3), disconnected hands, disconnected limbs, amputation, (semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime, doll, overexposed, photoshop, oversaturated:1.4)

The positive prompt involved was:

detailed (wrinkles, blemishes, folds, moles, viens, pores, skin imperfections:1.1), specular lighting, dslr, ultra quality, sharp focus, tack sharp, dof, film grain, centered, Fujifilm XT3, crystal clear

nutronicJul 25, 2023· 1 reaction

CivitAI

Does this work for version 5 of the model, its apparently also the last version.

mobilecaAug 27, 2023· 4 reactions

CivitAI

This negative embedding works wonders across so many models. Any updates planned? Thanks for the release.

eypachaAug 31, 2023

Yes. It can work better or worst but the most important thing its the base model. Other model just are fine-tuned versions of 1.5, sdxl, etc

darktensorSep 8, 2023· 2 reactions

CivitAI

Weighting this embeddings less than 1 (ex) 0.3/0.2...) will distort the picture strangely. Why does this happen when I expected the impact to be applied less because the weight is less?

mobilecaSep 9, 2023

I've also noticed timing it with [rv-n-e:0.3] for example causes strange artifact behavior too but I'm not sure if that's all negative embeddings or just this one

Rakly3Dec 8, 2023· 2 reactions

The embedding is exactly 75 tokens, adding weight to it exceeds it. Which will greatly impact the outcome. - (Ever noticed the numbers in the positive and negative prompts in A1111? They go x/75 , x/150 , x/225 , etc)

zerospamJun 27, 2024

CivitAI

Great work. Can this work any ethnicity - e.g. Indian?

TextualInversion

SD 1.5

by SeekerOfTheThicc

Download (Beta) View on CivitAI