Sign In

FLUX Has a "1Girl" problem (Quantization is making it worse?)

5
FLUX Has a "1Girl" problem (Quantization is making it worse?)

1Girl problem

In discussing a disturbing report I got back from a user of one of my LORA's with a moderator on this site, along with my own training of a cartoon LORA I have come across this information.

1girl in flux on a given seed will produce anime images, without the ability to specify young or loli in the negatives this is VERY problematic with realistic Lora's that remove clothing.

However simply putting down the "Ban Hammer" on 1girl in flux will cause issues with LORA's meant to produce anime, or toon based results, it could actually make those results worse.

Images

In fixed seed 500 NF4 Schnell we can see the results of 1girl, vs female.

(Note I think both of these images would cause issue)

1girl, face photo

female, face photo

VS the FP16 Dev version hosted onsite

female, face photo

What can be done?

From a site perspective FLUX is problematic to me without the ability to have negatives, but unique problems call for unique answers.

I am still unsure if Quantization + Lora is effecting the images in anyway.

It is a poor quality investigation to compare one seed but it does appear that quantization losses may cause smoothing that gets interpreted by UNET as make this person younger. (Or perhaps data was introduced into the quantized versions)

I have noticed some strange things with testing like turning the guidance up increases the time the image takes buy a large margin. (Maybe that is why Civitai has the guidance set at 3.5)

Prompt Importance

So in testing 18yo 30yo and 50yo don't seem to work, however thirty year old, and fifty year old do even when paired with 1girl.

If I have any suggestions to Civitai it would be to auto correct or possibly auto append:

18yo to eighteen year old

20yo to twenty year old

etc.

The downside would be that it is standard practice to finetune with 18yo, 20yo

5

Comments