This model is designed to help you build a character sheet for a character you plan to train, create references for generative video, or maybe create a 3D model... It generates the same character from multiple angles so you can isolate several images and expand your dataset. It’s far from perfect, Krea2 can already do something similar without a LoRA, with mixed results, but with this LoRA, using the triggerword (shown in the description) should reduce the number of generations needed. Just add the most precise description possible of your character.
For best results, generate in landscape format. You’ll need to find a balance: an image that’s too narrow may cut off part of the character, while one that’s too wide may produce more versions than necessary. Be sure to specify the background you want; I’ve deliberately avoided forcing a white background so you can choose.
There are some limitations. If the character isn’t described with enough precision, certain biases can appear: by default you’ll tend to get a mature Black woman, sometimes with duplicates. Very wide characters may overlap, and the different versions are relatively close together, so you’ll likely need minor editing if you want to isolate each version on its own image.
I hope you find this LoRA useful. Since it isn’t perfect, I don’t plan to charge for access. That’s why I’d really appreciate it if you could share your results below to help build a richer, more varied dataset for future improvements. I know I ask this every time, but so far, without responses, I’ve only had my own creations to refine my trainings.
If you like this LoRA, don't hesitate to:
publishing your photos directly in the model's gallery.
support me by giving some BUZZ (to train next models)
Thank you to everyone who supports me !
Let me know if SeaArt bot still publishing this model without my consent. My real LOREVER account is HERE with this ID: 7b81
Description
LoRA don't need to be too strength (0.6 can works to start) but it needs the Triggerword.
Trained on 2048x1363px
Keep neutral pose.
Be precise in your character description.
There are better models than the basic Krea2 checkpoint to make it works well.
FAQ
Comments (14)
I'll give it a try
Krea can already do this, I'm curious what this LoRA improves on?
As indicated in the description and the title, this helps you to generate a character sheet more easily thanks to the triggerword. However, nothing stops you from combining the two or continuing to do it without it. Just try it and see the difference, that's all I can tell you.
This LoRA saves the headache of having to prompt it in, and tokens for those paying for them but allowed to use LoRA. Why TF do people have to complain about sht without thinking about the answer themselves? I make character sheets all the time for use with Minimax H3 and this is a huge time saver, how could anyone not know this?!
Gave it a random try with a new character LORA, but it added e.g. huge tattoos all over the body.
Plain prompt copied from your sample images.
None of the training images contained any tattoos, makeup, accessories or weird hairstyle precisely so that you are free to add whatever you wish. If tattoos are being generated, it is probably due to either your choice of checkpoint or a LoRA you are using.
Maybe. I prompted the character nude. They never had tattoos anywhere.
Does it work Img2Img? If you already have a porttrait for example
It’s worth a look.
I never really used Krea2EDIT much, it didn't win me over when it first came out, but I’m curious to see your feedback on it, as it’s been a long time since I last tried it.
@LOREVER Aight...I am at training session now my self, but will give it a go later this evening :)
This is the most useful LoRA I've encountered for use with Krea 2 so far. This saves me a lot of typing and tokens.
Modifying the prompt a little we can also include a section for a facial closeup and some angles.
Thank you.
For those interested, you may want to try adding to your prompt for a more full character sheet, I put an example in the gallery and here's the prompt...
Candid amateur photography, slice of life.
Split image.
Left is a quad panel with front, profile and lips closed and parted.
On the right is the character sheet.
A Southeast Asian woman in bikini. White background.
My first training attempt actually used that exact layout; featuring a front and three-quarter close-up on the side, along with those three full-body views. But I had to change tack when I noticed the trained LoRA tended to generate the same close-up twice. So, if this LoRA does indeed allow for that kind of composition, I’m thrilled, because that layout was my initial goal.
In any case, thanks a lot for your enthusiasm; I’m always happy to see that my work has been useful to someone.
Metadata/prompts/checkpoint for example images?
The ComfyUI workflows are integrated in the pictures (with one visible character). Most of time I used the basic Krea2 checkpoint to generate my example pictures, and the prompt is simply the Triggerword first, with a small description of the character and I finished with the "white background".
It works easily with more descriptions, but the aim was to show that the LoRA is HELPING you to make a simple character-sheet.