CivArchive
    SameFace Fix [Flux Lora] - v0.1
    Preview 29786636
    Preview 29786681
    Preview 29786688
    Preview 29786715

    Important: This Lora should be used with a NEGATIVE weight.

    Also: The model is aimed at diversifying the look of women only. I plan to do a separate model for men later (and men with beards especially)

    Also cool fact: the size of the Lora is only 4mb (it is limited to 2 blocks with a high rank)

    • So, What's a "Sameface"?

    It's a common problem in generative AI, in which the model generates people with identical appearances. Unlike with SD1.5/SDXL, this is not just a result of the diverse dataset converging into a single token like "woman". in Flux, Dev/Schnell were distilled from the Pro version, resulting in a loss of diversity.

    • Why use Negative weight?

    Lora was trained on the most generic photos of women that Flux generates.

    If you then apply this Lora with a NEGATIVE weight, it will work as a sort of negative prompt, but without negative prompt!

    By using this clever trick, Lora will basically tell Flux to generate anything except a typical Flux face. And as result, we'll get beautiful unique faces every time!

    • Why Lora and not tricks like lower Guidance?

    People want diverse faces with high guidance in order to maintain high prompt adherence and image quality, that's why interventions at the latent space level are optimal.

    • How does it differ from other character Loras?

    While other Loras introduce their own "samefaces", this Lora, only tells Flux what should NOT be in the generation, not what to add. As a result, this Lora diversifies people's appearance, rather than narrowing it even further.

    • How was the Lora actually made?

    At first I was disappointed because nothing worked out. I was about to give up this project, but then tried to train Lora with just 2 layers, and suddenly during testing, I was met with beautiful faces, and with authentic, distant from Flux, appearance.

    • What's next?

    I have a lot to experiment with. Different learning rates, layers and ranks have to be tested in order to improve quality. Merging different Lora variants might fix the problems with artifacts at higher values. I'll make sure to make a better version, If this Lora has enough interest

    I'm really forward to your feedback, as this model is highly experimental. Any feedback is appreciated and will hopefully help me to improve the model.

    special thanks to Ostris for AI-Toolkit

    Description

    First Version!

    FAQ

    Comments (30)

    diogodSep 17, 2024· 1 reaction
    CivitAI

    Nice thanks! And I really hope you do a man variation. Using a low-rank (or just some blocks) is really smart here, since it won't learn a lot of other things to mess up the composition.

    trickybarrel72984Sep 17, 2024
    CivitAI

    Amazing results! Did you use ai toolkit and Which layers did you train?

    VadimFedenko
    Author
    Sep 17, 2024· 1 reaction

    Yeah, AI-Toolkit by Ostris. I trained layers 7 and 20, mentioned by Last Ben in this post x.com/__TheBen/status/1829554120270987740

    I'm going to try different layers though, to see if that yields better results for this Lora

    EbenezerDanglewoodSep 17, 2024· 1 reaction
    CivitAI

    Awesome!

    I like how it tones down the ridiculous blush and makes the faces more feminine.

    Hopefully the male version can remove beards.

    VadimFedenko
    Author
    Sep 17, 2024

    I used close ups of Flux butt-chins in the dataset, and it seems to improve the result. I'll be isolating the beards for the men version to see if that yields any improvements

    Mr_JingujiSep 17, 2024· 1 reaction
    CivitAI

    -0.6 and lower, lowers the quality of the image, can you find a way to fine tune it so it works better with Adetailer.

    Since this lora lowers the overall quality when using it in the main prompt I think I rather use it on the Adetailers prompt section so the rest of the image stays untouched.

    Think about it, it might gives better results.

    VadimFedenko
    Author
    Sep 17, 2024· 2 reactions

    I will try to balance the learning rate and experiment with merging to hopefully improve the image quality with this Lora

    groincSep 17, 2024
    CivitAI

    This Lora hasnt worked once. All I get is splodges of color like an abstract painting. I am using Forge though. Is there something I am missing? Thanks

    VadimFedenko
    Author
    Sep 18, 2024

    I did most of the testing in Forge and it was fine. Other people also report it works well. So not sure what can be the problem. Does generation work well without Lora/with other Loras? Splodges of color may indicate problems with sampling or Lora's weight calculation

    groincSep 18, 2024

    @VadimFedenko Thank you for reacing out. Ive only tried it with your lora, nothing else. I also notice that it says I have sameface2. But I notice the examples dont mention a version 2.

    skewbedSep 19, 2024· 5 reactions
    CivitAI

    Can you upload an fp16 version? I can't run fp8 loras on my Mac.

    Edit: I converted it myself on a Linux box.

    starzilla29377Sep 23, 2024· 2 reactions
    CivitAI

    This seems very useful in a lot of cases, but I have noticed, particularly I think it's with races that are a bit more unique to the system, the LoRA will make their faces really masculine. It will add very pronounced and manlike chins in some cases when I was working with dark skinned characters. In those cases, it seems like the LoRA is interpreted to mean the faces should be less women like and more manly. I am looking forward to you supporting men, because I think that may be at least a partial solution to this.

    SamLiuSep 28, 2024· 2 reactions
    CivitAI

    Worth a try !

    SendraelSep 29, 2024· 1 reaction
    CivitAI

    Works very good even at low values (0.3 or 0.4). Never use 1.0, that destroys the picture. Start with low values until you're satisfied. Haven't noticed much degredation on that values or interferance with other loras.

    FluffiesOct 1, 2024
    CivitAI

    Does it work for crowds of mixed men and women to make all of them look different together in the same photo when using a lora trained to resemble a specific person as the main subject?

    RefurrifyOct 9, 2024· 9 reactions
    CivitAI

    I'd love a version for SDXL

    DetroitArtDudeOct 10, 2024· 1 reaction
    CivitAI

    Using Forge I can't seem to get this to give me coherent results, but it could be me. With the weight positive, everything is fine, but negative tends to give me crazy stuff like hands coming out of random places, etc. Even at lower values like -0.2

    TyPorchDec 3, 2024
    CivitAI

    fp16 plz

    bissonfrederic69429Dec 23, 2024· 7 reactions
    CivitAI

    you should have named this version "-1" :)

    AnselmoMar 18, 2025

    -1girl (:

    humanppJan 1, 2025· 4 reactions
    CivitAI

    where men? :(

    FrameWorxJan 19, 2025· 1 reaction
    CivitAI

    This helps greatly but the strength you use it at depends on the model you use it with and whether you combine it with other loras. With dev nf4 v2, I find the setting needs to be lower to get away from that "flux woman" look (I'm so sick of that woman!). -0.1 to -0.4 I find she's still there, though not as obvious at -0.4. I've had better results with -0.6 to -1.4. The woman I did at 0.6 was pretty nice TBH! After that the image becomes faded and distorted. Some generations changed race around -1 but that's OK too. Thanks for making this lora. I'll see if I can figure out how to upload my image comparison with Flux NF4 v2 (I'm new to the site).

    hamstornadoJan 28, 2025
    CivitAI

    great lora! I was wondering if the same seed might result in different faces, but it stays the same - only different from the glossy catalogue style faces that Flux tends to do. thank you for this!

    LilyViciousMar 11, 2025
    CivitAI

    Hi there !
    That's a wonderful idea !
    Do you think your Lora would also work to fix the typical "flux-chin" or could it be the base for another Lora ?

    nick100Apr 14, 2025· 7 reactions
    CivitAI

    Flux has a huge problem with faces. It only draws one face. Well maybe one face of each race, even though sometimes different races have the same features too. If you could fix this that would be great. However so far I don't have much success with it. low values don't affect the face that much - slightly different, but has the same features. I had some success with -0.8 to -0.9 (or >-0.5). It still preserves some features, but it looks different. For example there are creases around the mouth and eyes that always stay the same. It does change the eyes which is some improvement. It's still a helpful lora, but I don't think it gives enough variety.

    txeexdJul 31, 2025
    CivitAI

    great lora

    HatolicAug 9, 2025· 1 reaction
    CivitAI

    This promt cure flux :)
    random angle view portrait random white young woman wearing random sport wears and hold random sport bottle in her hands, random poses in random location

    PS: latino or black or asian young woman

    slem_butterfly417Oct 13, 2025
    CivitAI

    Your idea of canceling something by putting a Lora in negative is a great concept, love it ! (at least i didn't know about it) Since then i try to negative everything and found out that by putting a nsfw Lora in -something put clothes on people keeping the silhouette given by "breasts loras". Thanks for that. 🙏🏻

    bigdog11Nov 5, 2025· 1 reaction
    CivitAI

    can you please make one for pony and IL , it's great

    carterscotthende7316Nov 9, 2025
    CivitAI

    It certainly fixes the Flux face problem, but it's mostly making unapealing masculine faces. Could the negative value possibly be making it remove feminine features entirely?