CivArchive
    🍒 Macromastia (Clothed) [Krea 2, Ideogram 4, Flux, Z-Image, etc.] - KLEIN 9B - V1
    NSFW

    Overview

    A clothed version of my other macromastia LoRA. This one here cannot produce X or XXX reliably.

    The model is pretty straight forward. Use the trigger early in your prompt. Further descriptions to the breasts position, framing and focus can help a lot to emphasize the effect.

    Sometimes it creates thicker or chubby bodies, for that you can include descriptors like slim or skinny or use negative prompts.

    Values over 1.0 may produce heavy artifacts. Decrease strength if you encounter artifacts. Most samples use a strength of 0.9.

    For more, have a look at the example images/videos.

    LTXV 2.3 (Sulphur 2) - V1

    Coming in June...

    Krea 2 Turbo - V1

    Use the trigger mcrmst and describe the breast/cleavage in some detail. Use the WAN 2.1 VAE for more realistic outcomes. Use the filter bypass LoRA for better results.

    mcrmst, A photograph of a woman with gigantic breasts, her voluminous, heavy breasts dramatically stretching and filling the soft beige cable-knit turtleneck sweater, creating deep curves and visible fabric tension despite the modest high neckline, sitting on a picnic blanket in a serene autumn forest.

    Ideogram 4 - V1

    Experimental model. Things break and body horror may occur.

    Trained on bounding boxes and JSON, but can be used without JSON structure prompts. Use the trigger gigantic breasts on a bounding box that should represent the woman's breasts. This way, the size and position can sort of be controlled.

    • Training model: Ideogram 4 fp8

    • LoRA strength conditional: 0.7-1.0

    • LoRA strength unconditional: 0.3-0.5

    • Sampler: euler / euler a

    • Steps: 20+

    • CFG: 3.0-6.0

    {...} // have a look at the showcase prompts

    WAN 2.2 T2V - V1

    After using the trigger macromastia, use additional triggers large, huge and gigantic to control the size further, see example video. Can be combined with Thiccification.

    • Training model: WAN 2.2 T2V

    • LoRA strength: 0.5-1.0

    • Sampler: euler / euler a

    • Use TripleKSampler with 1/4/4 steps or no lightning LoRAs at all for quality.

    • Use breast physics LoRAs to improve jiggle and motion for very large breasts. I highly recommend Huge Natural Breasts Physics at strength 0.3-0.5 for the high noise, no low noise needed or it will degrade texture quality.

    macromastia, a 21-year-old ginger woman drinking a milkshake from a straw in a cafe. she's wearing a tight blouse and jeans. She has gigantic breasts and a large cleavage. Her face is pretty and freckled, she has big lips. Looking out the window, sitting at a table.

    Flux.2 [klein] 9B - V2

    Use the trigger mcrmst and describe the breast/cleavage in some detail.

    mcrmst, A photograph of a woman with gigantic breasts, her voluminous, heavy breasts dramatically stretching and filling the soft beige cable-knit turtleneck sweater, creating deep curves and visible fabric tension despite the modest high neckline, sitting on a picnic blanket in a serene autumn forest.

    Z-Image - V1

    This model may also work with Z Image Turbo on higher strengths (1.0-2.0). You can also give Z Image Distilled a try.

    After using the trigger macromastia, use additional triggers large, huge and gigantic to control the size further.

    Experimenting with negative prompts and strengths can be very rewarding.

    • Training model: Z Image

    • LoRA strength: 0.3-1.0 (up to 2.0 on Turbo)

    • Sampler: res_2s+beta57 / euler / res_multistep / DPM++ 3M SDE

    • Steps: 40+ (8+ on Turbo)

    • CFG: 3.0-5.0 (1.0 on Turbo)

    macromastia, A high quality photo of a beautiful young woman with long flowing auburn hair standing gracefully as the central figure in a vast field of tall sunflowers during golden hour sunset. She wears a light, flowing white summer dress that catches the warm breeze and is stretched over her large breasts, her skin glowing in soft natural light, subtle smile on her face.

    Z-Image Turbo - V2

    Not compatible with Z Image Base.

    • Training model: Z Image De-Turbo

    • LoRA strength: 0.3-1.0

    • Sampler: euler / res_multistep / DPM++ 3M SDE

    • Steps: 8+

    • CFG: 1.0

    Prompt example:

    A photo of a cute woman with macromastia lying on a beach towel, wearing a yellow bikini revealing her breasts. Her breasts spill out the top of her bikini, forming a significant cleavage. She's bathing in the sun smiling at the camera. A summer day and warm candid setting.

    Description

    FAQ

    Comments (17)

    Griphen116Jan 18, 2026· 2 reactions
    CivitAI

    Can confirm, this does not work on 4b.
    Can you do one for Klein4B? It's easiest for picture editing and faster.

    cupra
    Author
    Jan 18, 2026· 2 reactions

    Too bad, but to be expected :(

    I may train a Klein 4B version tomorrow or the day after.

    Griphen116Jan 18, 2026· 1 reaction

    @cupra Naked version of 9b incoming though? 👀

    cupra
    Author
    Jan 18, 2026· 1 reaction

    @Griphen116 that is currently training, but honestly Flux.2 is fighting back a little. It doesn't understand anatomy as well as WAN or ZIT or even Flux.1 did. It may need some longer training or more images.

    hughchungus420Jan 18, 2026· 2 reactions

    @cupra It's funny how many lewd/near nsfw scenarios it seems i can generate with klein 9b out of the box, including some decent proportions, i'm not surprised its fighting back trying to train nudity but i just wonder what BFL was thinking here. Like they're dipping toes into just giving us total freedom without completely destroying their model for once.

    Anyway all that to say i'll be watching your career with great interest.

    Griphen116Jan 18, 2026· 3 reactions

    @cupra Yeah, I can feel the frustration from trying to use base.
    I with Z-Image Edit came sooner, because Klein just beats out anything for editing photos but sucks with nudity.

    hughchungus420Jan 18, 2026· 2 reactions

    @Griphen116 I'd even say Klein trades direct blows with z-image in every respect too. But its still beaten by the fact that the z-image team wasn't completely afraid of nudity lol.

    cupra
    Author
    Jan 19, 2026· 4 reactions

    @sneedingonmyligma420 thanks for the input! :)

    I don’t really get it either. It was pretty easy to train the clothed macromastia LoRA, but the naked one is just fighting back so hard. It looks like it doesn’t understand nudity well enough to handle extreme proportions.

    I will have to do some more experimenting.

    FightTheDeadll8765Apr 14, 2026

    @cupra is the 4B version available yet?

    cupra
    Author
    Apr 14, 2026

    @FightTheDeadll8765 Not yet. Most people I encountered used Z Image Turbo instead of Klein 4B, so I left it at Klein 9B. You think 4B is worth it?

    hughchungus420Apr 14, 2026

    @cupra Personally no lol. 9b is just better than the 4b all around, hardly a reason to train for it when few people use 9b as is.

    FightTheDeadll8765Apr 15, 2026· 3 reactions

    @cupra For potatos, yes.

    cupra
    Author
    Apr 15, 2026

    @FightTheDeadll8765 okay I’ll add it to the training queue when I’m back.

    green_angerJan 19, 2026· 4 reactions
    CivitAI

    For zit it's one of not so many working loras that do their job. Klein is only recently out and you've already created a working lora for it, super nice job! I especially appreciate you split the concept into two clothed and naked versions, very helpful and make it much easier to work with.

    With Klein I've problems with the base model, tried fp8 and gguf, the new workflow from comfy and from this lora examples (which I assume use the same), but base gives awful results, no detail, flat colors, most of the time not photo style at all. Once I switch to the distilled version, all is good in 6 steps, this lora works. What am I missing? All settings are the same except for steps and cfg.

    cupra
    Author
    Jan 19, 2026· 1 reaction

    Thanks!

    Yes this model got trained on the base model, but for inference it is quite useless. This LoRA has to be used with either

    https://huggingface.co/black-forest-labs/FLUX.2-klein-9B

    https://huggingface.co/black-forest-labs/FLUX.2-klein-9b-fp8

    in order to work. So the distilled versions.

    green_angerJan 19, 2026· 1 reaction

    I'm using the second one and it works nicely. The problem is the base version (fp8, same size) doesn't give any good result even when no lora is used. I mean this one:

    https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9b-fp8

    It's referred in the comfy Klein 9B TI template, but gives worse results than the second you linked.

    cupra
    Author
    Jan 20, 2026· 2 reactions

    @green_anger exactly. The base model is only good for training. It performs worse at extreme anatomy and has an overall worse image ELO rating. Sorry if I didn't make that clear enough.