Flux 1 Dev likeness of MaryLee, a fictional Asian American woman.
Notes after a week of use.
This is my first LoRA for Flux, and it is highly dependent on the same training I used for the same likeness for SD. As such, it does lean a bit into the plastics type look, especially at higher strengths.
This is also a relatively "weak" likeness model. Compared to my earlier versions in SD, it's a poor direct and consistent training, and acts more like a strong suggestion than a clone.
One weird thing about Flux this LoRA is good for.
FLUX 1 has an interesting habit of having maybe 100 or so likenesses in the base training. You see this over time, with the same few faces (shapes, features, expressions) being the default for a gender + ethnicity + age.
MaryLee v1 is luckily pretty similar to one of these base likeness, and seems to be consistent in triggering that same vague face.
Description
Updated training, with expanded references (80 vs. 20 in v1.0) and GPT-4o tagging based on updated community notes around Flux's tagging system.
FAQ
Comments (2)
V2 makes nice images. My only issue is that it tends more towards half-shots/cutting off at the waist than the first one. v1 was better about generating variety if the pose/shot was left unspecified.
Just realized that I'm an idiot and have been posting all my Flux content under a different login.
You can view my earlier SD model of this likenesses here: MaryLee for Stable Diffusion