Each “version” of this LoRA is actually a different LoRA with its own capabilities, but all of them are intended to improve the base Krea2 model in some way.
As I already did with SDXL, I believe this is a valid approach: using LoRAs to gradually improve the base model until it eventually becomes worth merging them into an entirely new model. At this stage, and considering that Krea2 is already quite good, my main goal with these LoRAs is to address what I see as its biggest weakness: photorealism, and especially the terrible “AI face” look that, in my opinion, still affects most generated faces.
The currently available LoRAs are GP (Glamour Photography) and BP (Beautiful People), each trained on a different dataset of carefully curated, high-quality photographs. Even though the training datasets focus on people (well… women, actually), their extremely high aesthetic quality makes these LoRAs useful for virtually any type of subject. The overall idea is simply to produce better-looking images.
There are no strict rules for using these LoRAs, since their purpose is broad rather than specific: to make images better. You can also combine them for a stronger effect (it is actually the best option for more realistic and impactful images), or experiment with strengths slightly above 1.0.
Take a look at the example images and comparison charts to get an idea of what these LoRAs can achieve.
Description
This is a Lora trained on ~700 high quality, carefully selected photographs with high aesthetic appeal in the "Glamour Photography" style.
FAQ
Comments (7)
Exciting to see some familiar faces. This looks really cool, worth disabling my lora stack to try. :)
Planning any new checkpoint for Krea2?
Nothing planned. For photorealism, Ideogram 4 is all I need and I don't see Krea 2 reaching that level of quality (but I may be wrong). But Krea is still fun to play with for everything else and Loras are easy to train, so for now I am just training some Loras :)
It looks like BP increases eye spacing and GP doesn't. Is that accurate?
I guess it is: BP has been trained on many high end portraits of really beautiful and unique women (think top-models). So PB indeed will impact more the face and push it to quite unique features ... eye spacing is one of them :)
@aramintastudio Okay. I remember a popular Flux "beauty improvement" checkpoint that showed a before and after with a man. The original Flux man had narrower eye spacing, which I thought was more attractive. But the wide eyes spacing trend does seem to be appreciated by many. It points to the subjectivity of beauty and may also be related to one's ethnic group. At least one study found that people tend to think their ethnic group's features are more attractive and the hypothesis was that it's due to familiarity, which also is interesting since, as you mentioned, some kinds of unusual features also get attention in the modeling world.
@ss9999 Well, tell me Anya Taylor-Joy is not attractive ... she has extremely wide eye spacing which is for me indeed very attractive because it gives her an otherworldly look :)
And do you tell me that you find Harry Melling attractive? Well, obviously beauty is in the eye of the beholder :)