Summary: --> Unlock Klein for nude females (NSFW)

Version 2.1 (Krea)

I'm happy with this one! Compared to Klein, Krea 2 has better prompt adherence, less body horror, and handles multiple characters better, but on the downside I get a bit more clone army in groups of people so it requires the prompt to be upsampled (prompt enhancement) to add some variety and not have the image be obviously "AI generated".

For inference you want to use Krea 2 Turbo (Raw is only for training) with guidance 0, steps 8, and no negative prompt.

Same guidelines as for Klein: if you get unwanted nudity turn the strength down to 0.8 or 0.7, or prompt a little more specifically with clothes for the non-nude people. With strength=1.0 and you say "nude girl on the left, and a girl on the right" it might make them both nude. You can say instead what that girl on the right is wearing.

Small note on Qwen Image 2.1 attempt

I tried to train for Qwen but couldn't get a good result. Qwen seems fussy. In contrast, Krea trained so easily. I tried many changes vs. my Klein setup such as lowering LR, lowering rank to 16, lowering alpha to 4, timestep weighted -> sigmoid, additional training resolutions. I aborted but can revisit if anyone has definitive ideas.

If you want to use Qwen try the Alpacas NSFW one, that's pretty good. But overall in my testing I'm getting better results than Qwen with Krea 2 and my LoRA (just released, see above).

BTW, Qwen Image 2.1 does reasonable nudity out of the box, if you fuss with guidance to like 2.5, steps 40, and put in the negative prompt "blurry, low quality, low resolution, artifacts, washed-out colors" to get good skin. It's a bit slow for me, though, and my results with the default model weren't extraordinary enough to want to use it a lot.

Version 2.1 (Klein)

This is a minor upgrade with:

• This is more aggressive LoRA (both 4B and 9B variants) so it has a lower threshold for triggering. It should generate nudes more easily, especially for prompts like "panties lowered to show pussy" or scenes where nudity is less expected like "college girl on lawn naked". If you get unwanted nudity or body horror set the strength down to 0.8 or 0.7.

• Soft dicks. If you want hard dicks and sex you still have to use SNOFS or UNCHAINED or similar, but then IMO you will get a more porny look. This LoRA should make girl-next-door beauties for you pretty well.

• I switched to ai-toolkit instead of SimpleTuner, which seems to have improved the quality. It's a bigger LoRA too for the same dimension but I didn't see any downsides in my regression testing.

Version 2.0 (Klein)

This is a major upgrade with a much bigger model, more data, and a fix to a SimpleTuner caption UX glitch (*see note below). I'm really happy with this one. It triggers more reliably, leaks less when you don't want it, listens better to boob adjectives, and makes pretty nipples. In short, this model fucks.

Okay, it doesn't really fuck, because it doesn't do dicks. Only female nudity. But it really rocks. This upgrade puts it on par with the best nudity LoRAs on Civit, or at least in the ballpark based on your personal preference. I made a post comparing 20 such LoRAs here.

Comparison spoiler: UNCHAINED is the undisputed king for just absofuckinglutely putting tits and pussies everywhere you ask for them, and even places where you don't, but it does give a bit of porn look IMO. FLUX 4 PLAY is another one in that category, triggering very reliably but with effects like huge labia. "Realistic Nudes" and "Perfect Average Females" and this Party Time v2.0 LoRA are next in line, I think, and behave similarly according to my testing, with different looks. I recommend taking them all for a spin on your prompts, it's free!

Trigger Words

This is a style LoRA with no specific trigger word. The qwen3 embedding is good at picking up any nudity as described in your prompt, so anything in the realm of "naked" "topless" "nude" "tits out" etc. should get the job done just fine and trigger the LoRA. Also it can infer nudity from context, so if you just say "woman in sauna" or "woman showering" she might show up nude even if you didn't say so explicitly, whereas in the raw (non-LoRA) model she'll often be wearing a towel or swimsuit with that kind of prompt.

Settings

For Klein9: set lora_strength=1.0. I use that almost all of the time. (UPDATE: I bragged too soon. I do have to set strength back to 0.7 for some nipple bleed-through and unwanted topless girls in some non-nude scene where you just happen to mention some girl's "breasts" or similar, which over-triggers the LoRA.)

For Klein4: set lora_strength to either 0.7 or 1.0. I had a harder time getting this LoRA to trigger in Klein4 (vs. Klein9) so I pushed the training a little harder. As a result, about half the time I get some nip bleed through for "thin shirt" (or similar) and back off the strength to 0.7.

Version 1.0 (Klein)

Verison 1.0 was a first attempt and I kept it lightweight (rank=16) with a "do no harm" philosophy. It's a decent LoRA but compared to V2 it doesn't trigger nearly as reliably or produce as nice looking tits and pussies.

Bottom line: Version 1.0 is OBSOLETE! Use version 2.0!

Technical Notes

This model works for inference in both distilled and base Klein. It was trained on base. I've switched to using only the distilled models now for inference since they are much faster and produce better realism.

For training the LoRA, SimpleTuner really nailed me with what I would call a UX bug! The default python I used (which came from ChatGPT, maybe it's its fault) had the setting for multi-line caption splitting in the wrong config.json file, not the data config, and nobody threw any error about it. I only found it by seeing the text cache log lines being the wrong number of entries, one for each line instead of one for each file. So for version 1.0 when I had multi-line captions like a line at the bottom adding "Overcast sky and natural lighting" in the training text-image pairs, it would split that and (randomly) give only that single line as the caption for the whole image in some batches. Doh! Effectively it was a form of extreme caption regularization. Fixing that made version 2.0 much better! (Oh, and a side effect, fixing this also nukes any issues with generating fuzzy or vintage images that was noted in Version 1.0 and probably was real, since the "low resolution" caption was getting discarded often in the training data.)