This is my first LoRa release, using 248 carefully selected images, iterating 6000 times, and taking 7 hours to train. It boasts amazing detail and generalization; it works very well. Feel free to use your imagination, and I hope you have fun!
Model Description: Surrealism, Fantasy Style
Trigger Word: kunge-fantasy
Suggested Weights: 0.8-1
Dataset: 248 images
Generative Model: krea2_turbo_int8_convrot
CFG: 1
Steps: 8
Sampler: euler_ancestral
Scheduler: ddim_uniform
Prompt Example:
A breathtaking surreal painting. In a dark sky studded with stars, a majestic angel kneels beneath the starlit night. The angel possesses enormous, exquisitely crafted wings adorned with shimmering patterns. She wears a flowing robe, reflecting the celestial light. In her hands, she holds a magnificent golden jug, from which a luminous liquid spills, cascading onto the vast, radiant earth below. This liquid, like stardust or divine light, spreads across the rolling hills, forests, and valleys, transforming the land into a dazzling tapestry of gold and silver. The angel's expression is serene and contemplative; her eyes slightly... closed. The painting is rendered in shimmering blue, gold, and green hues, with meticulous line drawing and striking contrasts enhancing its ethereal beauty.
Description
FAQ
Comments (4)
Its a gorgeous style. One of the best I've seen. I can't get an identical prompt, with the same settings to the same contrast and shininess. Are you using some post-processing on your samples?
All settings are in the model description. I later amplified them using seedvr2. Weights are set to 0.8, but you can also set it to 1; a weight of 1 will make the LORA effect more noticeable. Have fun!
Also, check if you've added the Trigger Word: kunge-fantasy?
Thanks for the response. I reckon the disparity might be SeedVR2. I can't reproduce that yet because I've got the old DiT weights so I'll get back to you after I've got the new fancy convrot 7b version downloaded.
By the way, I don't think including the trigger word makes any difference, at least not as far as I can tell.
Back: yes, it was SeedVR2. I've been able to recreate that look, cheers.









