Sign In

Make your own TI's with just a simple Prompt (Embedding Merge Tutorial)

Make your own TI's with just a simple Prompt (Embedding Merge Tutorial)

Preface:

This is a quick guide to making your own TI's. This does NOT REPLACE training them, there are many reasons training a TI is much better than doing it this way. But if you're looking for a way to store EASY prompts for future use? This tool is perfect. Also we're not sponsored by this.

If you want an easier way to read this with better language we're using GPT To re-word our articles and put them on our website! You can find this article here: https://www.end-media.org/post/the-easy-guide-to-creating-textual-inversions

EDIT: THIS NOW WORKS ON SDXL, there is an SDXL branch and instructions.

The GOD OF TOOLS:

Alright first of all take this : https://github.com/klimaleksus/stable-diffusion-webui-embedding-merge

Add it your extensions, I would hope by reading this you know how to install them. If NOT? Go to your extensions tab, please click on "INSTALL FROM URL" and paste that in the URL box, hit install. Then reload your UI after you're done.

(Please note, Batchlinks DOES support extensions)

How to make a TI

Cry.

Kidding, if you're unsure how the process of training one goes - don't ask us, we've only done it a couple times and panic at the timing. So with THIS tool you're basically creating a prompt in 75 tokens or less.


Your FIRST STEP after installing and restarting is the following:

Check your prompt in txt2img or img2img, make sure the spacing is correct and make sure it's under 75 tokens.

OK BUT STOP: ARENT PT FILES VIRUSES?

SIGHS - If you're making one yourself in A1111, you're not injecting any malicious code yourself, so for NOW IT SHOULD BE SAFE. It's not like you're scamming a twitch streamer or youtuber for a tesla scam...

There ARE PT to safetensors converters out there, and i'm not sure where.. so yea. I've used one but i lost the link.

MOST PEOPLE MISTAKE The converted files for LORAS because the practice of converting TI's to safetensors isn't nearly as big of a deal as CKPT and larger PT/PTH files.

Aight, move along now what do i do?

Clearly this can do MORE than just what i'm telling you it does, but I Don't understand vectors and frankly I barely passed high school and got lucky with college but if you tell me that 3x3 is 6 again i'm gonna slap you lol. (This was a rareinsults thing on reddit ignore me)

The prompt i told you to create, and check the token count? The text box that's larger, paste that in there.

The 2nd textbox you see - please make sure you have a RELEVANT token name for your TI or you'll never remember it. (Or be me, name one Enchilada-Remix on purpose so nobody knows you and a friend just literally made the most teenager level meme prompt to see what it would do).

Ok it failed.

You didn't check your spaces, or how long your prompt is.

Suggestion would be CUT THE FAT- you can add "OFFICIAL ART" Etc to your prompt later on, make your TI a subject, or in our case we've made meme's and song lyrics.

This IS NOT FOOLPROOF: TRAINING THEM DOES A MUCH BETTER JOB 90% OF THE TIME - Sailor Moon Goth was an utter FLUKE - only because the prompt has been used on SEVERAL MODELS across the last 3 months.

When you're making a PROMPT TI - make sure you know at least 5 models can handle the process. Not every model's gonna like OsenayanFace or TrumpMeme.

ADVANCED? Can you teach me?

Well no, the whole thing has multiple things and it's not the only plugin that works - but it so far as been the only one you can DIRECTLY JUST DROP WORDS IN. There's other ones you can use to directly edit vector weights.

Wait -- you made ALL OF YOUR TI's via this?

No, just 90% of them. Isometric dreams + NegEmbed are trained. As for the anime one we yeeted i don't know where that went, i'm gonna find it again and test it and maybe re publish or try it again.

Remember: 75 tokens. Spacing has to be perfect, and you can't do weights. SO you can't do it like a lora in your prompt. This won't work for prompting 10 loras. It's gotta be a generic prompt.

The 70s rockstar one?

Dan Avidan lycoris prompt.

The leotard prompt with it was literally Danny Sexbang from Ninja Sex Party lol.

Aight, well i'm exhausted.

I'm behind on AI because i'm in an insomniac phase and if i try and do AI. during my rocky reset sleep part - you'll get 2000 models in a week- nobody wants that off me.

SPONSORED LINKS

Our mixes are supporting PLURAL PRIDE MONTH AND DISABILITY PRIDE MONTH IN JULY.

We ourselves are not "A TEAM" we are a Diagnosed bodied individual with Dissociative Identity disorder. :3 No biggie, just don't start asking for 200+ alters in a list just cause you want different names of creators.

JULY IS PLURAL PRIDE MONTH - You all know who you are, and you shall fear no longer - you have space on CivitAI just as much as the rest of everyone else. Our goal is to create niche safe spaces for those like us. If you're not plural, neurodivergent - it's ok LOL - you're welcome to support and just download and enjoy our content!

If you want to learn more please go here: https://thepluralassociation.org/ and support us, because we're being fake claimed into oblivion for "not being ashamed".

Never be ashamed if you have quirks.

JOIN OUR DEVIANT ART GROUP: https://www.deviantart.com/diffusionai

Pre-Release Models, Lora Backups - Send us pizza! https://ko-fi.com/DUSKFALLcrew

ko-fi

WE ARE PROUDLY SPONSORED BY: https://www.piratediffusion.com/

If you got requests, or concerns, We're still looking for beta testers: JOIN THE DISCORD AND DEMAND THINGS OF US: https://discord.gg/5t2kYxt7An

Lora Request Form: https://forms.gle/NEkcMmVEjG4pBAP7A

JOIN OUR SUBREDDIT: https://www.reddit.com/r/earthndusk/

Listen to the music that we've made that goes with our art:
https://open.spotify.com/playlist/00R8x00YktB4u541imdSSf?si=b60d209385a74b38

BackMix List & Fave Models: https://rentry.co/duskfallcrew_fave_models_dump

We stream a lot of our testing on twitch: https://www.twitch.tv/duskfallcrew

WE ARE ON TENSOR ART NOW: https://tensor.art/u/611011406535381539

(*Tensor's links aren't working entierly, that's our profile)

23

Comments