The Easy Guide to Creating Textual Inversions (TIs) & Embeddings for Stable Diffusion via an easy Automatic 1111 extension!
Hey there! Looking for a quick way to store easy prompts for future use? Look no further! In this guide, we'll explore a tool that can help you create Textual Inversions (TIs) effortlessly. Just a heads up, this method doesn't replace proper training, but it's perfect for storing simple prompts.
Textual inversions, also known as "Embeds" or "Textual Inversions," are fantastic tools for prompting in Automatic 1111, especially in Stable Diffusion - Positive. They offer a way to create a unique style without completely changing the model's look. Trained Textual Inversions, or TIs, are equally important in both positive and negative prompt spaces, but they can be a bit tricky since they're trained on specific models. Some folks even call them "Positive" and "Negative" Embeds, fancy lingo for their different effects. The best part? TIs are compact, weighing in at under 1MB, while LORAs can be a hefty 50MB on Stable Diffusion. So, go ahead and explore the wonderful world of Textual Inversions – they're the cool kids in the prompt playground!
Oh, and by the way, we're not sponsored by this tool!
First let's get you a small image sampling of some of the wonders of the embeds we've made via this tool!
The Marvelous Tool:
Alright, let's dive right in! First things first, grab this marvelous tool from here: [insert GitHub link]. Install it as an extension (you know the drill!). If you need help with installation, simply go to your extensions tab, click on https://github.com/klimaleksus/stable-diffusion-webui-embedding-merge paste the link, hit install, and reload your UI. Piece of cake!
(Please note that the Batchlinks extension does support extensions, so if that's easier for you than using manual install I guess you can use that?)
How to Create a TI:
Ready for the fun part? Just kidding, but not really. If you're unsure about the training process, don't worry, we're not experts either. So with this tool, you'll be creating a prompt using 75 tokens or less.
Your First Step:
After installing and restarting, it's time to get started. Check your prompt using txt2img or img2img to ensure the spacing is correct and that it's under 75 tokens.
But Wait! Aren't PT Files Viruses?
Let's clear the air. If you're creating a TI in A1111, rest assured, you're not injecting any malicious code. So, for now, it should be safe. We're not here to scam anyone like those Tesla scammers on Twitch or YouTube, right?
There are PT to safetensors converters out there, but we can't recall where. We've used one before but misplaced the link. Just be aware that converting TIs to safetensors isn't as significant as dealing with CKPT and larger PT/PTH files.
Moving Along, What's Next?
This tool has many capabilities beyond what we're discussing here, but let's keep it simple. Paste the prompt we mentioned earlier into the larger text box you see.
For the second text box, make sure to give your TI a relevant token name, or you'll never remember it. Unless you're like us and intentionally name it something like "Enchilada-Remix" just for fun.
Oops, It Failed!
If your TI creation didn't go as planned, it's likely due to spacing or the prompt's length. Our suggestion? Trim the unnecessary bits. You can always add "OFFICIAL ART" or other details later. Make your TI concise and focused. We've had success with meme prompts and song lyrics.
This Isn't Foolproof:
Keep in mind that training TIs does a much better job 90% of the time. Sailor Moon Goth was a complete fluke, only working because the prompt had been used on several models over the last three months.
When creating a prompt TI, ensure that at least five models can handle the process. Not all models will respond well to "OsenayanFace" or "TrumpMeme."
Advanced? Can You Teach Me?
Unfortunately, we can't cover all the complexities of this method, and it's not the only plugin that works. However, it's the only one where you can directly drop words in. There are other plugins available to edit vector weights directly if you're feeling adventurous.
Wait, You Made All Your TIs with This?
Well, not all of them, but about 90% of our TIs were created using this tool. Some were trained with Isometric Dreams and NegEmbed. As for the anime one we mentioned, we misplaced it somewhere. We'll find it, test it, and maybe re-publish or try again.
Remember: 75 Tokens and Perfect Spacing:
To make this method work, keep your prompt within 75 tokens and ensure the spacing is just right. Unfortunately, you can't use weights like you would with LORAs. This method is more suitable for generic prompts.
For example, we've created TIs like a 70s rockstar prompt, a Dan Avidan Lycoris prompt, and even a leotard prompt featuring Danny Sexbang from Ninja Sex Party. The possibilities are endless!
SPONSORED LINKS
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
WE ARE PROUDLY SPONSORED BY: https://www.piratediffusion.com/
JOIN THE DISCORD AND DEMAND THINGS OF US: https://discord.gg/5t2kYxt7An
JOIN OUR SUBREDDIT: https://www.reddit.com/r/earthndusk/
WE ARE ON TENSOR ART NOW: https://tensor.art/u/611011406535381539
Comments