[flux2-klein-9b] Pyro's BJ
New Model, New Pyro
This is an experimental Lora
Research: How few ranks can you get away with?
Result: 90MB lora that understands more about anatomy than your biology teacher in 5th grade.
Also, it's literally the most trainable model I've ever trained with. It's so good I'm going to train it with 100k z-image-turbo images to make my own z-image-base because Alibooba is seemingly toying with our emotions. (If you want to sponsor this project shoot me a msg on discord)
How to use?
You can use this model with BASE and DISTILL. You can also edit images with it amazingly well.
Use BASE for better prompt adherence
Use DISTILL for images faster than you can jerk off
Pls try my distill workflow attached as TrainingData zip-File before telling me or the world that the distill doesn't work well. Thx.
Demo Images should contain workflow
In theory the LORA understands following concepts (in practice it's a bit shaky but that's why it's experimental, you'll figure it out):
General: a woman sucking a penis
Chipmunking or whatever this shit is called: a woman cheeking a penis
Deepthroat: a woman swallowing a penis
What could this be: a woman licking a penis
If you want her to look at you in disgust: pov
Will experiment further until I uncovered the mysteries of flux2-klein-9b and prompt adherence is at a point where you trust the LORA more than your mom.
Discord: https://discord.gg/PSa9EMEMCe
Technicals:
Framework: my diffusion-pipe fork: https://github.com/pyros-projects/diffusion-pipe
Dataset: 300imgs split evenly over all sub-concepts
Training:
65 epochs with 512p (AdamW - batch 4 - lr 2e-4)
15 epochs with 768p (Prodigy Schedulefree)
Version History:
bj_20260120_22-22-29epoch15_comfy: Initial Release
Description
FAQ
Comments (47)
Love you text to explain how this lora works!
is very good!
10/10 description format. Will download later and try.
Are you training locally (what hardware) or renting?
This works wonders.
Works well in edit mode, but text-to-image using the LoRA doesn’t work out of the box (it’s even worse when adding a character LoRA). I’ve only tested it with the distilled model so far; I’ll do more testing with the base model.
all images in the album are made with the distill so I can't confirm that it doesn't work well with the distill... pls try the workflow from those images and report back
I attached my workflow as zip if that makes it easier
What are you talking about? The LORA works flawlessly with t2I and I2I. You are doing something wrong. I get good gens 95% of the time.
I am getting very good results with both T2I and edit. Tried with two different Character lora and still getting very good results.
In case you hadn’t noticed, it greatly improves the quality of skin textures. Using the same system (I don’t know whether you used image pairs or just individual images for each concept), you could create a LoRA to improve skin realism.
Can you please tell how you gather dataset and what typical pair of images look like (or upload a few)? I want to understand how to keep good resemblance, quality. and flexibility when training lora.
freaking awesome
please make more nsfw loras for flux klein your loras are freakin solid
Amazing!
Can't wait to try this lora since the others I've tried for F2K9B have been surprisingly great. This model has so much potential. Thanks for sharing.
Wow!!!!! Thats just amazing!!!! Please dude dont stop make more!
can't get deepthroat/swallowing to work but other than that it does work.
Prompt a small, short, thick penis. It will look like deepthroat. I've also had some success telling it to have the subject sucking/kissing the crotch, and leaving a description of the penis out.
How to make woman's face to stay the same with image workflow? I have a problem that her appearance is quite different from provided image.
Yeah same, although still a good lora, depending on your goals.
This guy openly admits to not captioning his datasets correctly on Reddit lol. That's almost certainly the cause of that issue.
@ZootAllures9111 wait seriously?
@ZootAllures9111 couldn't find him saying that on Reddit. Even if, you doing something wrong bc this works pretty good most of the time.
If you use the NSFW and head_swap loras it works great also the anatomy slider can help if needed!
@Lora_Addict He made a post both when the original Flux came out and recently after Klein came out basically insisting that captioning very little is the "right" thing to do. Which is precisely what causes the sort of issue @guy33 was describing, on every single model architecture ever released. Anything that isn't explicitly captioned WILL randomly show up in generations and / or be impossible to change / remove in a way you can't control by any means beyond the actual lora inference strength level in ComfyUI or wherever.
@ZootAllures9111 yeah possible, the Lora still works amazingly good :D
I'm going to treat you all to something nice. Take a POV photo of your own cock, looking down at it.
Use this image as the only reference image with this LORA. Ask it to edit in a person sucking your cock, describe the person. Ask it not to change your penis.
Viola! Custom fantasy blowjobs.
it does not work, I tried to use Dredd's dick, it will always add dick that LoRa has been trained with, u have to use two images edit mode and then it will replace it but it will not look good as it just does copy paste over from one image to another - it does not know to make it seamless and match colors and textures, it treats penis the same way it would treat shoes, bags, watches...
@erdelman73267 0.6 strength is your friend and proper prompting
@erdelman73267 It does work and the prompting is simple. I turned around in my computer chair, snapped a picture, popped it into the image node in comfy. Then I prompted "A woman is kneeling, sucking my penis. POV Blowjob."
95% success rate.
@Drakeni and how to do the reverse? taken a picture of a woman and want to add a penis from a man, could be either pov or just Add it close to her face, or any other part. Extremely greatful for any directions, not found much that worked for me.
Thanks man, works great. Any way to get a black cock?
lol the people downvoting options they don't have to use if they don't want to. bless their terrified little hearts. must suck being in perpetual fight or flight mode
That is just gift from gods! Works so perfectly good. Any plans on other concepts?
I give it about 3 days before this gets yeeted. Great job tho!
I'm not up2date about civitai drama, since i basically never visit this site except when uploading shit.
Why would it get yeeted? Are there policy changes I don't know?
@pyros_sd_models I hope it doesn't, but a lot of similar Wan 2.2 loras (like the jump cut lora, and @ffjggrtbjibv's genitals lora) have been yeeted these past few months. Basically any lora that can turn an SFW initial image into an NSFW photo/video. But that's why we have Civarchive.
@anonz11 thanks for the heads up. what's the preferred alternative for hosting loras then? huggingface?
@anonz11 Civarchive? I cannot help but cautiously ask what (and where) is it? Tried searching, got messy different results. And I feel like it may be very helpful in the foreseeable future. A simple hint will do, if full address is not appreciated around here (who knows considering your own bet).
@green_anger civarchive dot com.
@pyros_sd_models yes, huggingface is what's mainly being used for mirror links.
Can this be generalized to any penetration act? I'm having fun making anime to realistic gens and most of the time the result is awful whenever a penis is involved.
probably not, this is a blowjob lora
i added this lora and penis lora and prompted
"POV of asian girl spreading her legs in front of a erect long penis"
"POV of asian girl spreading her butt cheeks while looking back at camera in front of a erect long penis"
"POV of asian girl in front of a erect long penis, she is spreading her butt cheeks while looking back at camera"
it took few iteration of run ( i always do batch run of 5-10) among them you will always find one which has accurate anatomy no weird shapes.
reason why i am using 2 different lora for it is because BJ lora works best with POV view.
penis lora that i got for flux adds very well accuracy in anatomy. keep 0.5 0.5 strength for both.
my man writing a prose in description
Very good model, it can be a bit sloppy at times (no pun intended), but overall it's great
Is it possible to make the testicles larger and the penis less veiny? Thank you.
f9 edit gave me horror lol
Details
Files
bj_20260120_22-22-29epoch15_comfy.safetensors
Mirrors
bj_20260120_22-22-29epoch15_comfy.safetensors
bj_20260120_22-22-29epoch15_comfy.safetensors
KLEIN_BlowjobPyro.safetensors
bj_20260120_22-22-29epoch15_comfy.safetensors
bj_20260120_22-22-29epoch15_comfy.safetensors
bj_20260120_22-22-29epoch15_comfy.safetensors
bj_20260120_22-22-29epoch15_comfy.safetensors
bj_20260120_22-22-29epoch15_comfy.safetensors
bj_20260120_22-22-29epoch15_comfy.safetensors
34234_bj_fk.safetensors
bj_20260120_22-22-29epoch15_comfy.safetensors
bj_20260120_22-22-29epoch15_comfy.safetensors
pyro_bj.safetensors
flux2klein_bj.safetensors
bj_20260120_22-22-29epoch15_comfy.safetensors
bj_20260120_22-22-29epoch15_comfy.safetensors
KLEIN_BlowjobPyro.safetensors
bj_20260120_22-22-29epoch15_comfy-mid_2326122-vid_2616688.safetensors
Pyro's BJ - Klein9B - pov,licking,cheeking,swallowing,sucking.safetensors
Pyro's BJ - Klein9B - pov,licking,cheeking,swallowing,sucking.safetensors
pyro_bj_v1.safetensors
bj_20260120_22-22-29epoch15_comfy-mid_2326122-vid_2616688.safetensors
10-bj_20260120_22-22-29epoch15_comfy.safetensors
bj_20260120_22-22-29epoch15_comfy.safetensors












