Transform drawing to more or less photorealistic image. The main difference from similar loras is that the dataset mostly featured girls with a European face type.
Description
This version usually renders the eye color correctly, without always trying to direct the gaze at the camera. The negative is that the clothes look a little plastic, they don't understand closed eyes, they can draw nipples where they shouldn't, they don't draw toes well.
FAQ
Comments (7)
This is surprisingly versatile and retains the details of the original image really well. Even worked on pixel art in my tests. Maybe the most interesting thing is that it really does make "cosplay"; I tried it on art of a succubus and the end result looked like the wings and horns were plush and 'hand-made'. I look forward to more of your LoRAs!
This looks great! If you want some feedback; if you can get the gaze to not change, meaning where they look, that would be amazing for concept art! <3
This is the 8th version, which is partially trained on the results of previous versions, including qwen, and I started from scratch a couple of times. The main problem was the departure from the purely Asian facial type in girls, which has been more or less solved. The second problem is the color of the eyes, which is mostly correct now. The third problem is the direction of the gaze, as I manually moved the pupils in the images, but this is not enough. Either they try to look at the camera or at the direction of the head. I want to improve it myself, but for now, I'm taking a break. I also forgot about the closed eyes, although I added winks. Once I rest, my goal is to add more options for open mouths, angry expressions, and emotions in general.
This is a really interesting Lora. But I'm not too keen on the blurring that appears on the skin, especially around the eyes. I tested it on mid-range shots.
thats a good main difference :D gonna try. (all other high strenght will produce asians)
It seems that there is a lot of Asian cosplay in the qwen and flux dataset - they try to generate it themselves. I haven't been able to get rid of it completely yet, it will be stronger on certain characters, weaker on others.
nice work...
if you use comfy you can downscale the input image by 2 but leaves the original path resolution ... you get a more realistic image