Tensor.Art
Create

3QQ is an Aasin women

LORA
Original

Updated:

3QQ is an Asian

learn how to train lora with flux ... 3 qq aim for over fitting ..yes

some people said its okay for overfitting and change the strength lora to 0.7

so i want to try his statement ,,make an overgfitting lora

i trained and tested on this website,, i cant test it on my gt 730 - -"

i have 10+ tested images now ,, i pretty sure the images tell the story enough

this is what i reviews

1) 3qq trigger dont affect thing , dont need to type it

2 ) this lora believe in prompt .but

  • -just use the word ,,,,, asian women , young asian womnen ,or

  • (a young Asian woman with a fair complexion and straight, shoulder-length brown hair with subtle highlights. Her hair has a soft, natural texture and is parted slightly to the side, framing her face. She has a delicate, oval-shaped face with high cheekbones, a small nose, and full lips with a subtle pink gloss. Her eyes are almond-shaped and a light brown color, accentuated with subtle makeup including eyeliner and mascara.) ( skin color , make up can change too i believe )

so many problem here such as

1) some images in my dataset make problem here .. i think i will remove her or dont double repeat on this lovely images ... i forget about " the earing " too

2 this lora ,, she come with earing ,,, i think i have check lots but there is image that she wear earing and i dont tag it ,,,

this is my data setting

36 image , all of them are low quality ,ai creation, its sharp 120 - ( 100 ,90 , 50 )

Base ModelFLUX.1 - dev-fp8

Trigger words 3qq Repeat7 Epoch16

Save Every N Epochs1

Seed-Clip Skip1 Text Encoder learning rate 0.00001

Unet learning rate0.0001LR

Schedulercosine_with_restarts

OptimizerAdamW8b

Network Dim64 Network Alpha32 Gradient Accumulation Steps1 Shuffle caption-Keep n tokens3 Noise offset0.03 Multires noise discount0.1 Multires noise iterations10conv_dim-conv_alpha-Batch Size-

Prompt 3qq , A solo girl with a beautiful face, sweet smile, and black hair with bangs, looking directly at the viewer. Perfect lighting and natural skin texture, soft shadows, and high contrast. Realistic eyes and well-defined facial

featuresSamplereuler

f

Version Detail

FLUX.1
3000
16
dont use trigger . this lora dont follow trigger ,, test it 3 time just use the prompt your want to see.. train with flux , aim for overfitting lora . asian character mess a lot ,, i use 500 / imag e number for epo number i will use last epo for this model ( epo 16 ) i will upload and test 15 14 13 12 11 10 9 8 ,,as i aim for the learning anyway .. the training is good enough (for me ) at epo 8 .. some of the charactore in my dataset popup , she dont look like a girl in the original flux .. anyway all of my dataset are a.i. with blure and not shape ( shape at 100 90 60 50 and so on ) welcome for suggestion and testing .. let s have some fun

Project Permissions

    Use Permissions

  • Use in TENSOR Online

  • As a online training base model on TENSOR

  • Use without crediting me

  • Share merges of this model

  • Use different permissions on merges

    Commercial Use

  • Sell generated contents

  • Use on generation services

  • Sell this model or merges

Related Posts