DISCLAIMER: THIS IS A CHARACTER MODEL, SOME PEOPLE ARE MISTAKING IT FOR THE NEW XLABS REALISM MODEL, THIS IS JUST DESIGNED TO CREATE THE WOMAN IN THE PICTURES.
This was just an experiment in character creation from a single image, this is not a real person and it is not intended to represent anyone living or dead. I also want to mention I am not associated with X Labs and their AI work, this is just for fun.
The big man himself, Maxfield, posted this model from X Labs which featured an image that became pretty controversial, in that a lot of people were questioning weather it was actual AI or not. The realism model itself is pretty controversial as some are claiming it doesn't really do anything, or it in fact makes Flux worse. I personally think it does work to an extent, but the differences are not necessarily better, I think it just adds a grain that conveys a minuscule amount of realism, almost like an Instagram filter.
To complicate matters further, some are claiming the image was created with the Pro model of Flux, and it doesn't even use the Lora. The original image was posted by Ted Chang found here,
This was mainly me testing flux's abilities to make characters from a very limited dataset. I use a few Photoshop tricks + InstantID to make the dataset for this character, and honestly the results are actually pretty decent. I also did some testing with and without the Realism lora from Xlabs which I have put in the samples as well as a comparison.
Weight for this is anywhere from 0.6 -1.5
No triggers for this, but its a good idea to put prompts like asian, dark hair, basically describing her features. She tends to be smiling a lot, so be prepared for endless smiles.
FOLLOW ME FOR MORE FUN EXPERIMENTS, TAKE CARE LOVE YOU BYE
Description
FAQ
Comments (14)
The original picture looks very real, the model you posted looks 100% AI generated version of it, so cannot fool people. If that's what you were going for, well done! I myself prefer realistic generations since it's a pathway to re-create dead people to one day resurrect and entertain us again.
I have no idea what point you are trying to make, or what you came to say.
I tried to make a criticism into a positive comment. I'll be more to the point, and it may come off as rude.
1) To me, this model generates fake looking people, it looks less realistic than Xlabs'. Look at the original picture's girl's arm. It has more detail. It's not just grain. This one generates a super smooth arm. No untouched-up woman's photo has such smooth arms.
2) However, many people like smooth skin tones with more contrast and fake looking people, so this model may be better to other people. Also, I think in comparisons to other fake looking models, this one is more realistic than most, but just not Xlabs or PixelWave level realistic.
I missed the part of the description that this wasn't a replacement for Xlabs model, it's just a character Lora based on 1 image. That's impressive.
@afterate2 Yeah its literally in the title of the model.
I think you also may have missed the part where the picture in the xlabs lora probably wasn't even created with that xlabs Lora or even with standard Flux, and on top of that the original poster does not leave a prompt for that image, making replication pretty difficult, but not impossible.
You missed a lot of points and came across a bit condescending to be honest.
I actually think mine looks pretty good, if you actually take the image and zoom in to the arm as you said it still has a lot of detail. Mine has a lot more color and vibrancy and the skin in some areas looks a bit more shiny, but to the untrained eye its still passable as a photo.
I also show you a comparison of images with the different the loras applied and also mixed in. Clearly showing the difference. You kinda made it seem like I was trying to hide something, which is why I had no idea what you were talking about.
If I wanted to I could easily apply Photoshop filters with various grains to make it more believable for you. But I was just trying to create a character Lora and I leave my prompts so people can replicate my work, I also give instructions. I put in a lot of effort a lot of people don't appreciate but they are quick to jump the gun and criticize. This was not intended to win a realism contest. So unfortunately with this you are way off the mark.
@Wolf_Systems I understand your thoughts on my criticism.
As a final comment, I want to say that I don't have yellow buzz to give, but I gave a like to your model and to some of your images as appreciation. I know it takes a lot of effort and time to create and gather the data, train a model and pick the best epoch, and to respond to my messages.
So you took 1 picture, cropped it in various ways and ed it to FLUX and it was this good?
Something like that, I changed the background, applied different filters, changed her clothing, but I also used instantId for sdxl to generate new images for the dataset, then I altered and cropped those as well.
@Wolf_Systems hmmm. What happens if you just feed a single picture as your dataset to flux?
In the past inised flux to create a datset to train a pony model. But i had like 10 low quality pictures to train flux
@barlogspam176 You cant feed a single image, I mean technically you could, but it would be a terribly inflexible model, as in you will have very little ability to change anything. Even this model with the dataset I created is not the most flexible, and I could probably make it better by running it through the trainer a second or third time with the new images I generated.
so, if I understand correctly, it only takes one image and the workflow you described to generate a well-trained lora model. Congratulations! I like it a lot. I think i want to replicate your work! :-)
the model's flexibility is still fairly limited because it comes from 1 altered image. So it's not as strong as the usual character models .Well trained and flexibility dont always go hand in hand.
Any chance you could share the workflow you used to get the images to make this? I'm having trouble finding one and FLUX pulid doesn't work well for me, atleast not for men. The ones for sdxl i was actually able to get to work didnt generate good faces and i couldnt figure out how to add a face detailer
Not sure what you need, the workflow is in the image, you can just download the image then drag n drop into comfyui, but this is a flux model, not sdxl, it will only work with a flux checkpoint.