CivArchive
    Female - Face Portraits - Detailed Skin - Closeup Macro - Flux - v1.0
    NSFW
    Preview 44816171
    Preview 44447797
    Preview 44434938
    Preview 44434993
    Preview 44435012
    Preview 44437476
    Preview 44450316
    Preview 44457313
    Preview 44454324
    Preview 44624999
    Preview 44806053
    Preview 44806071
    Preview 44806102
    Preview 44807095
    Preview 45176444
    Preview 45350276
    Preview 45516744

    Trained on 200+ very highly detailed high resolution (min 2000px) images of amateur female face portraits, with visible skin pores, skin hairs and blemishes/moles. Multi resolution trained with crops to 1280px on tiny details.

    Before releasing it here I merged with 5% Female Tongue, Mouth and Teeth Lora so it also still able to produce good looking full body shots etc. as well. But the main focus is on face only portraits.

    Use descriptive prompts like:

    • A closeup photograph on the little skinhairs on the chin of a young woman

    • A photograph of a young amateur woman's face with brown eyes, long hair, and her mouth open. The focus is on her mouth, emphasizing her perfect symmetric teeth. The image is detailed, capturing the natural detailed texture and pores and blemishes and hairs of her skin and the subtle sheen on the inside of the mouth.

    Guidance / (distilled) CFG between 1,6 and 2,2, Important to keep it low for realism!

    Check the photos in the gallery, they are all by DPM++ 2M sampler (so nothing exotic) and all have prompts and generation data to inspect.

    (note: some of the sample prompts mention a different lora filename (tongue-v3) I could not get rid of that after deciding to publish this lora as its own closeup model and not a version update, its the same lora as you download here)

    Description

    FAQ

    Comments (8)

    sevenof9247Dec 12, 2024
    CivitAI

    looks nice, maybe you can reduce the rank and than ~100MB ;)

    and btw , you can leave the brackets, no different ^^

    jhnprst2150
    Author
    Dec 12, 2024

    thanks i will experiment with lower ranks 16 or 8 for next version. not sure about the brackets you mean in the sample image prompts? they do differ for me in forge (i.e. lower the weight of the words)

    sevenof9247Dec 13, 2024

    @jhnprst2150 maybe its 32bit not 16 ???

    and yes the brackets in prompt, for me in forge it does change only a very very minor bit (delte all brackets vs none)

    ptdurnvcoedqegsdnwDec 13, 2024

    Forge (and comfy iirc) don't have working prompt weighting for Flux. SDNext claims to, but I haven't been able to verify. So the brackets probably aren't helping, they're just acting as a token and adding minor noise to the conditioning. But sometimes a placebo effect is real so who cares?

    jhnprst2150
    Author
    Dec 13, 2024· 1 reaction

    @ptdurnvcoedqegsdnw wow i had not thought about that really, thanks for the comment! i'll check it out

    jhnprst2150
    Author
    Dec 13, 2024

    @sevenof9247 16 or 32 bit do you know how i would check it? I used bf16 precision during training and save? the blowup in size maybe because I used svd_merge_lora.py from sd-scripts before release., the original file before merge was half the size.. maybe something happened there. for sure next release I will take note of the filesize, its indeed huge,..

    sevenof9247Dec 14, 2024· 1 reaction

    @jhnprst2150 in Kohya GUI you can choose save precision too, the script i dont know.

    AlentejoJan 4, 2025· 1 reaction

    @jhnprst2150 The current closeupface-v1.safetensors is float32, if you are still wondering. I think rank 32 is very adequate and I wouldn't be bent to change it, but you should try to retain the original bf16 data type.