- This topic has 10 replies, 4 voices, and was last updated 1 year ago by
.
-
Posts
-
Well, since I hear this question from time to time, is it possible to get a consistent character with consistent clothes in stable diffusion, even while using LORAS?
Simple answer...YES!
HOW TO:
1. Generate an image of your character with SD and note its SEED.
2. Enter the SEED in the SEED entry (just in case you were generating multiple images).
3. Activate ControlNet and choose as pre-processor "reference_only" and put the image you generated in the image slot there.
4. Now you can generate your character with different LORAS, in different positions, expressions, etc. Backgrounds are not consistent, though.REFERENCE PIC: (no LORAS)
IMAGES WITH VARIOUS LORAS (EXPRESSIONS)
What a great tip! Thanks.
I wonder what some of the other pre-processors do. There's dozens of them.@Frank22 Thanks, I am sharing info as I learn it. The problem I am finding is that even if someone explains the various pre-processors of ControlNet in a guide, I usually cannot make heads or tails about what they are talking about, or find their info pretty useless.
Example:
The Preprocessor reference_only is an unusual type of Preprocessor which does not require any Control model, but guides diffusion directly using the source image as a reference.
This can be used to make images of a similar style, especially anime and cartoons!
Would you have guessed that it can be used to make consistent characters even with loras? Me neither 😅
Thankfully, I was watching a very unrelated video about text to speech and someone mentioned this technique.
If you want to read the guide, good luck 🤣
@Frank22 I found something new, maybe this one can be of help:
Thanks... but I'm still on 1.5. So many new things at Civitai are now SDXL I might have to jump soon.
OK... I jumped. Just DL the XL base, refiner and VAE. Presently DLing RealVisXL V3.0. It worries me that some of the XL Loras are big files, scary big like over 1GB each 😲.
I hold you personally responsible for my ever diminishing disk-space Legolas.Still so many girls with 6 fingers....
I consider daz clearly better when you need consistency (but will probably not last)@Corbeau There is just 1 image with 6 fingers 😑
@Frank22 😂 Don't forget to download the ControlNet models for sdxl!Here are a few recommended LORAS to "Fix" things that have worked well for me:
https://civitai.com/models/200251/feet-xl-10-sd-15
https://civitai.com/models/200255/hands-xl-sd-15Also, I recommend using sdxl-vae-fp16-fix (it works better for me).
@legolas18
I don't want to argue and it does not matter as I was speaking from a more general point of view but I count 2.... (images 1 and... 5 two thumbs for the right hand). This AI bias will be cured at some stage (as people with 3 tits or 2 anuses...).
Thanks for the interesting fix links. I will try a new test probably (my grail would be from my Daz renders to create animated series...).
Merry creative Christmas !@Legolas18
Thanks for once again throwing out some interesting and useful info.
These will be cut & paste to my AI for idiots notebook.
🙂P.S. @Corbeau
Who the hell is paying attention to their fingers ???It takes me roughly an hour to train a Lora with 4 epochs and 30ish images with my 4070. I see some at Civitai saying they train with hundreds of epochs. Unless they have some sort of trick with their Kohya settings, that must take an awfully long time even if epochs/training time isn't strictly linear.
- You must be logged in to reply to this topic.
