Person Enhancer
Works on SDXL, PONY, and ILLUSTRIOUS models.
Person Enhancer helps with images of people (male and female). It works best on images of a single person, but can also be used on groups. It helps refine facial features and to a lesser extent bodies. It also helps with details in general and giving images a more high definition aesthetic. The sample images are from a variety of checkpoints including at least one image from PONY, ILLUSTRIOUS, and SDXL.
It is designed to help with these items - it doesn't do miracles. If a model is bad with hands or eyes, this may help but it will not fix a flawed model.
It was trained on a diverse set of images so it should not try to over impose it's will on your base prompting unless it is extremely overweighted.
Suggested weight is 0.65-0.90. It can vary based on model. Different models and model types impact the optimal strength so some trial and error is needed. It works for both men and women.
V2 - USE THIS VERSION.
V2 is a major upgrade over V1. It is a little stronger too so I dialed the recommended weight down a little. Sample images all done at 0.85, but optimal can vary by checkpoint. Note this is a "flux look" so if too strong it will give you "flux chin and mouth". A version may be released without the flux bias in the future.
Disclaimer: The user assumes any and all liability for the images they produce. The creator of this LoRA is providing it for educational and ethical purposes. It is up to the user to abide by all laws and ethical considerations in the jurisdiction in which they reside.
Description
V2 - Major upgrade. V2 was trained over IllustriousXL instead of over SDXL like V1. It yields good results over Illustrious, NOOBAI, PONY, and SDXL.
FAQ
Comments (4)
For stability and style enhancement for lora's characters: realistic background,realistic girl,realistic skin,photo realism,photorealistic
This looks potentially useful to me, as I've been trying to squeeze semi-real looks out of non-realistic illustrious models. Was this trained entirely on flux-generated photorealistic images? How large was your dataset?
V2 was trained with just over 1700 images. It is roughly 50/50 between FLUX generated images and hidef real photos of people.
I have used all three version (0.7, 1.0, 2.0) extensively with different checkpoints. 2.0 is overall the best version but has a strong limitation making it unfortunately unusable for me.
When I come from a non-western face, the Lora turns it into a western face and then keeps it stable whatever I do to my prompt, even when adding certain types of Lora that usually tend to change the face. This is a really rare behaviour.
The main reason why I prefer the old Beta 0.7. is that is works across all kind of faces, even when it is mostly useful on Pony SDXL checkpoint. But even a change from a real checkpoint to a 3D checkpoints the face is still recognizable.
So overall one of the best face Lora's that I'm aware of.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.