FINALLY!
I managed to create a nel TI that i'm happy with and that delivers consistently good results. This was trained on NAI with 4 different data sets on multiple different weights over 35,000 steps with over 400 images.
I know that by now there's an amazing LORA by Rakosz but i prefer TI's and really wanted to make it work.
This is a textual inversion embedding for Nelliel Tu Odelschwanck, or short Nel, from Bleach. Use vslinx_nel to activate.
Recommended is to use Clip Skip 2 since the embedding was trained on it!
Feedback is very welcomed and i'd love to see your creations with it if you'd like to attach them in a review.
The images were created with different models:
Description
FAQ
Comments (6)
Looks good!
Any plans on Nemu, Rangiku and Nanao?
Currently working on a Lisa, Nemu is next on the list since there’s already a great Rangiku Lora out there, will definitely do her at some point but she’s further down the list, added nanao to my todo list as well, thanks for the recommendation!
Whats the difference of making it as a TI instead of a normal LoRA?
Mainly the way it's used. You have to include the LorA either with Additional Networks or with the Standard <lora: etc. and then you have to set a weight etc.
With a textual inversion you can change the name for example to just "nel" and now when doing your normal prompting the TI will automatically loaded when you use the word "Nel" in your prompts. So it's trigger word sensitive. Obviously you can have different variations inside the TI just like in the lora depending on the tags of your training images. For example you can add "hair ornament" on my inoue orihime TI in either positive or negative prompts depending on if you'd like them included.
Lastly to get a good TI sadly takes a lot more time. You need a good amount of images (around 150) to get somewhat good results and the training takes way longer :/ usually for 10000 steps it takes around 8 hours.
Hope this could clear some stuff up 🤝🏽🙌🏽
May I ask, big brother, which large model should I choose to approximate your displayed image?😁
Hey my friend,
the first few images were all created with the BloodNightOrangeMix, the ones i added later (with higher saturiation) were all created with Corneos 7th heaven mix.
BloodNightOrangeMix : https://huggingface.co/WarriorMama777/OrangeMixs/tree/main/Models/BloodOrangeMix
Corneo's 7th heaven Mix :