The Lora is trained on 100 images I generated using Midjourney 7. All images were captioned with joycaption-beta-en (natural language + English). Personally I'd rather not caption the images but I had to in order to proceed with the training. I've used lots of Midjourney loras, with various architectures over the years and to be honest they're not great at replicating the Midjourney aesthetic. However, I think this one is the closest yet. Maybe it's just the magic of Z-Image Base. All the images I used were realistic so maybe not a good idea for generating Anime or art. The showcase images are in pairs, the first image of each pair used the Lora. Enjoy!
Description
This is trained on over 100 images I made using Midjourney v7. If you are looking at the training details, it is not a mistake, I found the first epoch (20 repeats) to be the best. Not sure if there is a point making a Midjourney v8 lora as I'm not too impressed by it. This should work on Krea 2 Raw also, as this is what it was trained on. The first set of image use the lora.