Old School Runescape (OSRS) style lora for Flux 2 Klein 9B.
Reccomended Strength 0.8-1.0
Tested on distilled model both for Text2Image and Image edit style transfer using the default ComfyUI workflows. Generated with 6 steps and res_2s sampler.
Model was trained on base model using AI-Toolkit. All training imgs were taken from screenshots and character models of OSRS.
Old School Runescape (OSRS) style lora for Z-Image Turbo.
Reccomended Strength 0.8-1.0
First try at Z-Image lora. Works most of the time, but has a chance to have some (severe) artifacts depending on the seed.
Please feel free to share your thoughts and images!
Description
Flux 2 Klein 9B version for OSRS style. Tested and works for both Text2Image and Image Edit style transfer.
For style transfer simply use: "Transform the style of the image to OSRS style."
Trigger Word: OSRS style
FAQ
Comments (7)
Works GREAT
Works nicely, i have a question tho, how many images did you use for the training, i'm new to the whole flux 2 klein image editing part, i've only done some LoRas for Qwen image
I used 70 images of random screenshots/character models and captioned them with simple natural language myself with no mention of style.
I included a trigger word to make it easier to transform other images with the lora. The training configuration was the default on AI-Toolkit expect for differential guidance of 3 and 3500 steps total.
I can attach a zip of my dataset and training config if you would like.
@ShadyNyaa Yeah, that’d be great actually. If you don’t mind attaching the zip with the dataset and config, I’m curious to dig into how you set it up.
@Mascarade14 No problem. I attached the training data. Hope it helps.
If you find anything else interesting on training loras, feel free to share. I am always on the lookout to learn more.
@ShadyNyaa you are a life saver it worked perfectly, of course kinda expensive but i think that by my next run i will increase the learning rate a little and decrease the amount of steps to a 2800 or 3000, cuz it already works well, maybe more images into the mix like 120 max to push the model a little. (Do note that i used RunComfy for the training since my pc would have taken a week otherwise)
@Mascarade14 Glad it all worked out! I have yet to find the perfect sweetspot for images to steps, but so far Klein is pretty easy and fun to train, especially for styles.