Fast Negative Embedding (+ FastNegativeV2) - v1.0 Strong

NSFW

Fast Negative Embedding

Do you like what I do? Consider supporting me on Patreon 🅿️ or feel free to buy me a coffee ☕

Token mix of my usual negative embedding. This way it's faster to use and easier to reproduce.

I'm also working on a much stronger one that doesn't preserve styles but it's good on base models without style loras applied. I plan to upload it during the next week, so be sure to leave a ❤️ to be notified of future updates.

~~This is very much a~~ ~~work in progress~~~~, but since I'm already using it in examples, I thought I should upload it.~~

It should keep styles on all my style loras. If not please tell me in the comments.
Additionally, if you find this too overpowering, use it with weight, like (FastNegativeEmbedding:0.9).

Update: added FastNegativeV2. It shouldn't be necessary to lower the weight.

Of course, don't use this in the positive prompt.
This includes Nerf's Negative Hand embedding.

Description

This will alter the style a bit but tries to greatly improve the final result. Might not work on all models.

FAQ

Comments (21)

coffeebeansMay 21, 2023

CivitAI

Only for Anime?

Lykon

Author

May 21, 2023

it should work on NED too as it's basically my usual negative prompt converted to a single word. I'll make sure to test it accurately and post some examples. Check back in a couple of hours

AptronymMay 24, 2023

@Lykon How do you make something like this? Turning a whole prompt into a TI?

Lykon

Author

May 24, 2023· 1 reaction

@Foulman there are tools for it, like Embedding Inspector, but not limited to that

TealonMay 21, 2023· 5 reactions

CivitAI

Thank you for the work! Would you happen to have the safetensor format for the model?

Lykon

Author

May 21, 2023· 2 reactions

pt should be fine for embeddings. If you're worried you can use an inspector, but I can assure you this is safe :)

nutronicMay 23, 2023· 1 reaction

CivitAI

When enabled it puts into my negative prompt box as <lora:FastNegativeEmbedding:1>

Is that correct or should I remove the < > ?

Thanks

Lykon

Author

May 23, 2023· 14 reactions

hi @nutronic , this is not a lora, but an embedding. So you should just put it into your embeddings folder and simply type the filename in the negative prompt. Like in the examples basically. Just plain without "<>" or anything else.

To change the weight you can do like "(FastNegativeEmbedding:0.9)", to keep it at normal strength, just "FastNegativeEmbedding" (without quotes)

13127834121318Jul 7, 2023

@Lykon thx

A1322May 31, 2023· 15 reactions

CivitAI

I threw it in the positive prompt expecting eldritch horrors and it's all circuit boards and what looks like 70s-era computer parts. It's kinda cool, actually.

Lykon

Author

May 31, 2023· 1 reaction

interesting. What about the strong one?

A1322Jun 1, 2023· 3 reactions

It appears I have been duped into entering the Jaws of Hell. Bad-Prompt is scarier but the "You. Shouldn't. Be. Here." energy here is off the charts.

Anyway, now I know how TI training works. Thank you for the spook.

VendaciousnessJun 1, 2023· 3 reactions

Welp, I know what I'm doing, today... I love technology stuff!

A1322Jun 1, 2023· 2 reactions

@Vendaciousness the Strong version is much spookier, but if you start clip skipping it, it turns into a creepypasta for (fake) unsettling DOS games from like 1990.

118121Jun 7, 2023

CivitAI

Does this affect the overall art style of the image?

Lykon

Author

Jun 7, 2023· 1 reaction

Are we talking about photorealistic images, anime or semi realistic? THis will affect realism, it's not indicated for that. It's good on semi realism and anime. At 0.9 weight won't impact the style. I use it all the time.

LostAtSeaJun 17, 2023

@Lykon when do you personally use this vs easynegative and when do you use them both together

Lykon

Author

Jun 17, 2023

@LostAtSea I never use easynegative, has too much effect on styles. Plus I think it's trained on asian stuff or something, because it sometimes changes blonde hair to black hair.

118121Jun 18, 2023

@Lykon anime and 2.5D models such as perfectworld or kotosmix

LostAtSeaJun 18, 2023

@Lykon "Plus I think it's trained on asian stuff" That might explain why I always find the faces to be very similar regardless of model. Can I get your got to negative embedding copypasta? This is what I usually rock "easynegative, bad-hands-5, verybadimagenegative_v1.3, (low quality, worst quality:1.4), bad anatomy, bad composition, sketch, greyscale, monochrome, ugly"

Lykon

Author

Jun 18, 2023· 8 reactions

@LostAtSea I usually go FastNegativeV2 alone for anime/art or BadDream, (UnrealisticDream:1.2) for realistic stuff

TextualInversion

SD 1.5

by Lykon

Download (Beta) View on CivitAI