Trained on hand selected 130 images. Might make a larger version with thousands of images. I wanted to avoid this becoming something like OpenJourney which looks nothing like Midjourney V3. I want this LoRA alone to be able to surpass Midjourney V3, not merely replicate or lag behind it. That’s my goal.
I’m happy with how it turned out, it’s much more coherent than midjourney V3 yet preserves its style as long as your prompt stays within 1-3 tags.
I might release a future version trained on more images so that it can still apply the style on longer tags.
Recommended settings to maximise style:
Lora Strength: 2
CFG: 2
Sampler: DPM ++ SDE Karras
Resolution can go as low as 256×256 as the original images were of this size. While the training set is predominantly 256x, 512x works just fine, still sharp and all. 768x Also works. Experimentation is recommended.
Something I also found out is that, this LoRA is trained on base 1.5 and base 1.5 isn’t very coherent. Midjourney V3 is even less coherent. But if you use the LoRA with a model like dreamshapers 8, not only is it the beautiful MJV3 style but its also coherent.
So this model is even better than midjourney V3.
Legal stuff:
[Midjourney Trademark] is a trademark of Midjourney, Inc. We are not endorsed by or affiliated with Midjourney, Inc. We hope “Style LoRA” makes it obvious that this is indeed an addon to Stable diffusion and not in any way, shape, or form, a product of, or affiliated with, the Midjourney, Inc (the corporation.)
Outputs of AI image generators’ copyrightability depends on jurisdiction and usually on the human who was prompting the AI (if copyright can be applied at all), no images in the training set involves work of humans whom I have no permissions to.