Type | |
Stats | 396 0 |
Reviews | (58) |
Published | Aug 24, 2023 |
Base Model | |
Trigger Words | 4WBK |
Hash | AutoV2 8B3E6E8586 |
Hey there, if you're creating tons of pictures daily for family and clients, this type of LORA could be your new friend.
I want to speed up my workflow and also test SDXL’s limit so I decide to train 4 DIFFERENT ART STYLES using only 1 LORA, and to push it further I mashed them all into a single 1 collage picture. Which mean this LORA could rock 4 different Art Styles in just 1 image.
SETTINGS
WEIGHT: 0.8-1 to make a collage of 4 images. Or 0.4-0.7 to blend all 4 styles which will give birth to the 5th original style. (Image 10, 11, 15 & 17)
FOCUS: It's trained to draw PORTRAIT (human) - my primary use case. If there's demand, I'll expand this Lora's capabilities to include full body, animals, landscapes, etc.
PROMPT: You can allow it to draw freely (5th image), or for greater similarity among the four images, include specific description (white hair, pink crop top, etc.)
OTHER: Compatible with Auto1111 default setting, let me know if you got good results with other settings
TRIGGER WORD: 4WBK (optional)
MY MAIN GOALS:
SAVE MOTHER EARTH: Also cut my bill & spare my old GPU
4X FASTER BRAINSTORMING PROCESS: No 1024x1024 needed during brainstorming, Quadruple the speed is more appetizing.
SIMPLIFY TESTING PROCESS: Skip XYZ setup & repetitive style name typing. Plus keep a XYZ slot free for other purposes
DESTROY THE WALL BETWEEN REALITY & ART STYLES: Mix it with a REAL FACE LORA & enjoy (may work with Character Loras if they could draw portrait). Also I don’t like it when some LORA doesn’t draw male or other things well so using Combo Style LORA may help with that.
CREATE NEW ART STYLE: An unexpected bonus I found out after training >> Setting weight at 0.4-0.7 could blend all 4 styles which will give birth to the 5th original style.
INPAINTING: I got some very interesting results. However, the result is a bit random due to the fact that SDXL is not an inpainting-focus model and this LORA only train on 28 images. I expected that we could solve this with more training
I only have enough time to create 112 close-up portraits for training. I'm not sure, but I think it needs at least 400 images to establish a stable connection between 4 styles (full body, animals, architecture, etc.).
FUTURE PLANS:
If there's enough interest, I'll generate more training data and retrain the model.
If you need it, I can also create four Lora models for the four styles I used here.
In the future, I might increase the total number of styles to test the limit of SDXL
If you like it please:
Upload your images on CivitAI so I can join the tournament too. There is only a few days left and I need 50 images from you guys to join.
Give me suggestions to further enhance this Lora. You con comment here or DM my Discord
Again, this is just the first version, I hope more people will enjoy this & we'll see more LORAs of this type down the road.
Cheers!