CivArchive
    Faye - v1.0 - Drift City
    NSFW
    Preview 142474870
    Preview 142374615
    Preview 142374614
    Preview 142374618
    Preview 142374617
    Preview 142374608
    Preview 142374605
    Preview 142374610
    Preview 142374611
    Preview 142374613
    Preview 142374620
    Preview 142374609

    Faye is trained from 100% generated images! Any likeness to person(s) living or dead is as coincidental as finding your own doppelganger out there - could happen, not intentional. Her reference images were created using Krea 2, Flux 2 Klein, and Gemini. Then she was put through a bunch of tuning passes.

    She is done... Final is a purely re-trained version built from generating about 2,000 images to get proper proportions, angles, shape, and frame out of the multiple failed models - of which, I narrowed to 21 diverse photos to capture as much range as possible. I couldn't leave the problem unfinished. She is complete.

    Sick Ollie, Analogue Madness, and anything from Stable Yogi as a base model can drift her age young in some cases, just like her other attempts. The best I can attribute this to is hair, face, and neckline tapping into blocks that are of younger women - she meets the same metrics as Tegan and Carys (and they don't drift young in the same prompts).

    Likely will have small lapses of look, as I couldn't get EVERY image to be completely identical to character. Drift you see in the previews is mostly base-model drift (I have tested her in 12 different base models and she mostly carries her look throughout).


    For context, I now have about 10 different scripts to extract spatial, textural, and compositional data about characters, 8 different compression and pruning scripts, and 6 re-weighting and mergine/combination scripts. You can't easily fix a poorly trained model! Your work is best suited for building CLEAN and SOLID training data! These tools will (and have already) helped build out more generic LoRA for you all. I hope to release some of them as I get them more broadly tested.

    Description

    FAQ

    Comments (18)

    Otto_MattiqSep 10, 2026· 1 reaction
    CivitAI

    I attest that Faye is indeed flexible and fits in.

    xFennec777
    Author
    Sep 10, 2026· 1 reaction

    Oh my. What have you done to her so soon?!

    LurkingAISep 11, 2026· 1 reaction
    CivitAI

    Can't wait to test this - esp. after your Carys model turned out so amazing. So many LORAs, so little time...

    xFennec777
    Author
    Sep 11, 2026· 1 reaction

    Let me know what you think. This one has a few flaws, but seems to hold a great character base that responds well to correction.

    I am thankful that the influx is slowing down. A little. A few weeks back there were so many models getting uploaded each day that I couldn’t even keep up looking at the preview page!

    LurkingAISep 12, 2026

    @xFennec777 Hmm okay, so i spent literally the whole day (12+ hrs) trying to get this to work. But it is super-inconsistent for me (using Sick Ollie checkpoint).

    1. Character / Face - it shows a generally very different character than your previews; both your Turbo previews but also the Into Realism previews. But then, this result also changes radically depending on what else is prompted. Sometimes she looks like a 3/10 below-average woman, sometimes she has this stunning perfect grin with a triangular chin/jaw section and gentle cheekfat swell and dimples (think Angela White, but better) which you want to prompt the whole day. But it's impossible to get a consistent result, and it's a lot of other factors to blame. All in all, i guess it always shows one of at least 3 different people depending on circumstances (like facial expression, or anything listed below).

    2. Body - this changes so wildly it isn't fun. Sometimes she has tiny pointy tits but a grotesquely huge ass, sometimes a pleasingly compact body closer to your previews, and then again a (natural) voluptuous body with a perfect soft curve and round breasts. Mind you this is all without even mentioning her body in the prompt. At either rate, the woman in my results weighs at least 10 extra pounds and is often closer to 18 than 29.

    3. Clothes / Accessories - Have a massive impact on her face and body. Let's say a crop top gives her big round tits with nipples peeking through, while a swimsuit makes her athletic like a olympia winner. Besides that, i have certain prompts which work about 95% with other character LORAs. But with Faye, it plainly ignores that at least half the time. E.g. oversized Navigator sunglasses are randomly Wayfarers or Ski Goggles (monolens); black heels are suddenly beige; a plain coca-cola shirt has extra nonsense print.

    4. Posing - Model is often stubborn about assuming a pose (albeit facial expression worked quite well and expressive). I frequently got amputated legs and extra hands in POV sex scene attempts.

    5. Locations and environment - state e.g. "Havana" which should put the scene into the typical colonial streets with beautiful lights and colors, but not with this model - she's either indoors, or (if you prompt an outdoors place) - it's all grey and desaturated.

    6. Age - this is the most bizarre one. Prompting e.g. "eats a popsicle" and she looks like 15-16. Wears a cardigan and Gen-Z eyeglasses? chubby 17-year-old without a chin. But then you state an age (e.g. 29), and she loses all the details in a different image which she would have without stating an age (e.g. wrinkles from laughing).

    7. Hair / Eye color - not stating a hair color gives her random hair colors (light brown to light blond). Eye color was actually green in my first prompt (should be blue according to your previews?). Well, if you mention these as dark blond and blue eyes in the prompt, the face changes and these headparts are larger.

    8. Style / Composition impact - is fairly huge, and often in a poor amateur way (unless the stars align and a specifc shot looks great). Often forces a strange close-up high-angle shot, or directly does selfies without mentioning that, or makes the character nude by default. And where the Carys LORA allowed for deep shadows and smoothly flowing light/highlights, it often looks dull and uniform with the Faye LORA.

    So besides trying to force a specific look by adding age, hair/eye color, even body height and figure type, i also tried to increase the LORA strength by up to 1.4. But i just can't get her to look like a single person between two prompts. And that's before having all the other problems.

    What's also strange is - when i try any of the prompts where Carys looked so natural and lively/lovely, this character looks dull, stiff, boring. On top of that, the image style also looks poor.

    xFennec777
    Author
    Sep 12, 2026· 1 reaction

    @LurkingAI Super helpful! I had spent a lot of the day getting similar results - the character drift is huge in some models BUT much more consistent in others. Fineporn had similar findings as sick Ollie - mega drift. I wonder if I have created a drift checking monster. Base Krea 2 has some clothing/scene/pose dependent drift, but she keeps the same generally look (similar for Solstice and Into Realism, though i think she conforms to the base model’s core woman which changes her a bit).

    what it did feel like was running a well trained character LoRA at like 0.6 weight in some models. This is why the age shift - some of the models are trained hard on younger women.

    If you prompted “blonde woman” in place of “Faye” and dropped her LoRA - do you get similar age/weight shifts? If so, she make be too malleable and I dropped too much of the core character (or she may play better at a 1.5 strength, meaning I need to reweight her).

    xFennec777
    Author
    Sep 12, 2026· 1 reaction

    @LurkingAI Will spend more time cooking. Cary’s and Tegan are the go to for now. Hope I can get Faye right!

    SUPER helpful! I appreciate you!

    LurkingAISep 12, 2026

    @xFennec777 Well, just glad to give some actual feedback.

    As mentioned, i also used prompts i had in use for e.g. Carys (several 1:1 reprompts with just Faye instead of Carys LORA and trigger) but it vastly differs. Like that one shot she exits the white BMW, Faye suddenly looks barely 18, and a totally different person than that Navy combat suit shot. One pic Faye was riding a Harley in full leather suit, looked like legit 12 just based on body proportions and the face under the sunglasses. And that was with only Faye as prompt, no mention of age.

    The closest thing i got was one candid bikini shot with the LORA at 1.4 STR, but this ruins the rest of the image (e.g. a previously razor-sharply rendered popsicle was a compressed mess of blur) so i didn't upload that.

    And as i said, in the shots where she looked good, she looked really good. I liked those the most where she looks like a soft-curvy 25-33yo with the big yellow / violet tinted eyeglasses (black triangle bikini / red-white coca-cola outfit). Though i think you were going for the more lithe-looking type.

    xFennec777
    Author
    Sep 12, 2026· 1 reaction

    @LurkingAI I found some potential sources of the wild drifting. One of the LoRA I merged in for proportion adjustments has a lot of face data in it and it seems to have had an exponential effect here. Still getting drift in latest tests, but significantly lower range.

    Hoping to find a few more fixes over the next day or so. Appreciate the added findings! I think certain prompts tapped that area of the merged LoRA and carried extra data through.

    xFennec777
    Author
    Sep 13, 2026· 2 reactions

    @LurkingAI I have stripped blended and retrained, I see the same issues with age/popsicle, sweater/glasses, wrinkles depending on the prompt (but mostly in Sick Ollie and Fine Porn). Into Realism and Krea 2 Turbo both have a TINY shift on those prompts. I wonder if there is something about the body/face that the core Krea 2 tunings lean age heavily on this one. I MIGHT have a more solid base, but her unprompted age is more like 21 not 28-32 like I intended. Will see what I can find.

    LurkingAISep 13, 2026· 2 reactions

    @xFennec777 Good to hear. I also loaded up Tegan last night, also gives some very good renders. My GPU is just really really slow (1m20s to 2-3mins per image) so it always takes me at least a day to review a character LORA.

    xFennec777
    Author
    Sep 13, 2026· 1 reaction

    @LurkingAI It takes me 40s or more per render. I may try to work in the 4step LoRA to speed up (but that’s more testing on character modification)

    xFennec777
    Author
    Sep 13, 2026· 1 reaction
    CivitAI

    Faye currently drifts a lot. Sorry, and thank you for all the feedback. Her training seems to have bad data somewhere in it - I have yet to figure out what bit. Which breaks all the other modifications I normally do (and so amplifies the drift). I will crack it!

    LurkingAISep 15, 2026· 1 reaction

    Spent some hours testing. I counted at least 5, maybe 6 different actually distinguishable people being displayed. It doesn't even keep the same hair color between two shots, even if you prompt it.

    Renders look more crisp, at least.

    LurkingAISep 15, 2026· 1 reaction

    Okay so i just uploaded this one set with 20 bikini posing pics at a tropical waterfall. She looks fairly consistent there, but i think it's partially the tinted glasses and the specific combo given for age (29), hair color, hair style, and the bikini.

    I think the set looks absolutely gorgeous. The images have "soul" if AI can have that. Render quality is also great this time.

    xFennec777
    Author
    Sep 15, 2026· 1 reaction

    @LurkingAI I think this data set is cursed.

    I scoured it, re generated a few bits. then regenerated - it didn’t lock in body/face/texture cleanly at the same point, and the tools that stabilised Tegan is what caused the silly 1.0 model here.

    For 2.0:

    I wrote scripts to extract and enforce facial structure and body structure.

    I even did some testing of these tools on Carys to see if the introduce drift… nothing, she rendered perfect.

    This data set trained well in Flux 2 Klein :/

    LurkingAISep 15, 2026· 1 reaction

    @xFennec777 You could check in with SickOllie, i heard he has some good tips about character LORA creation. As i personally have never bothered making one.

    xFennec777
    Author
    Sep 15, 2026· 1 reaction

    @LurkingAI thanks. May reach, but this is the first one that has failed so badly (and oddly that it trained well in another model). Kinda want to just toss the set :)

    I have a ton of tools built - like I can get a 95% likeness of Carys in a 9mb LoRA.

    LORA
    Krea 2

    Details

    Downloads
    231
    Platform
    CivitAI
    Platform Status
    Available
    Created
    9/10/2026
    Updated
    9/30/2026
    Deleted
    -
    Trigger Words:
    Faye

    Files

    FayeV1.safetensors

    Mirrors

    CivitAI (1 mirrors)