Harmony in Fusion is a LoRA that gracefully merges traditional Eastern aesthetics with contemporary digital creativity, creating art that exudes balance and serenity. Trained on artistic images featuring yin-yang symbolism, this model captures the essence of harmonious contrast, weaving subtle and sometimes bold yin-yang shapes into each piece. Every image is an infusion of complementary elements, from koi fish in synchronized flow to tigers and dragons locked in symbolic duality. The delicate dance of light and dark, warm and cool, conveys a sense of unity amidst opposition.
This LoRA doesn’t just create yin-yang art; it infuses it with a digital twist, bringing “Harmony in Fusion” to life—both as a balanced composition and a harmonious infusion of traditional symbolism into modern form. Use this LoRA to create oppositional subjects that flow harmoniously in balance.
I've debated releasing Version 3.0 for a few months. Rather than having to prompt a yin-yang shape, the model pulls your image in that direction regardless of whether or not you prompt it. I think the effect can be fun so I decided to share it. Version 2.0 is still the more flexible one, but Version 3.0 can lead to some interesting art.
Usage
To use the most recent version of the LoRA, use the following settings:
Trigger word: ink and brushstroke diyngyng style, as "a diyngyng style watercolor painting" or "a diyngyng style pencil sketch."
Other tokens that work well: yin-yang symbol, yin-yang configuration, contrasting shades, opposing, opposite, intertwined, circular
Lora Strength: A strength between 0.8 and 1.2 is recommended; higher strength should enhance the abstract quality of the style.
Description
FAQ
Comments (9)
What checkpoint did you use? When using FluxDev my images with your prompts, a cfg of 3.5 and ~35 steps all look like "pixelated".
Any suggestions? Thank you.
These were all generated from the base Flux Dev 1 f8 version, available for download on Civitai. They were trained on the same model (I just checked the metadata to make sure it is "flux-1-dev") and at 1024 resolution. The only difference I see from what you shared is that I very rarely use more than 20 steps.
I'm seeing the same situation, it seems like it was trained on a low res or compressed dataset and is imitating the compression artifacts as a style. I love the concept but the Lora itself seems to be a dud.
The sample images the author has posted have the same issue to.
@Walternate It's likely a bit overtrained. To get the yin-yang shaping to work, it had to be a bit overfit, so the model tended to force things in that direction. It was trained from a single image, but I have a larger dataset. Honestly, it was an experiment, and I wasn't going to revisit it, but I'm pretty sure I can make a better version.
@Dark_infinity I'd say go for it, because the concept does have potential and could be fun to play with!
@Walternate I've got it ready to go. If it goes well, I'll have an update tomorrow.
@Walternate I'm hoping this one turns out better -- at least for the image quality. I've got a few more ideas on how to make the effect stronger without over-training, but see if version 2.0 works out better for you.
@Dark_infinity just tested it out and it works amazing! a massive improvement over v1, well done!!
@Walternate Thank you. I appreciate the feedback on both versions. If I don't know anything is wrong, I can't fix it, and my testing doesn't always catch everything. Most of the time, if I have version 1 posted for a model, it's often a beta that seems to work well during testing, so I don't mind constructive critiques. Thank you again; without help and feedback, I couldn't have made it better.