NEW VERSION V2
Trained with higher res images. (Z-Image only)
Much better pussies (but still not as good as I would like).
Changed the tags. Use the phrase
"faceforwardlookback kneeling pose"
"faceforwardlookback standing pose"
"faceforwardlookback lying pose"
combine with anything else you want.
Z-Image is good at poses anyway, and genitals exist, but they look bad. This aims to improve it.
Makes an image of a woman facing away from the camera, looking back at the viewer while showing her naked pussy.
Made without faces in the training data, so shouldn't alter your existing faces too much. Using 2 Loras will drop the strength of the other Lora however. Try dropping the strength of this Lora as far as you can to maintain the face.
Description
First version
FAQ
Comments (21)
awesome job work even with other loras
"Made without faces in the training data, so shouldn't alter your existing faces"
For some who dabbles here and there in training loras, could you provide example what you did. Did you just cut/blur the face in the dateset? Going to try this out. When using characters loras or even controlnet in comfyui to get this pose is not easy without it getting messed up.
It was an experiment that seemed to work.
This one uses all 512x512 images. All with that look back pose (or near enough)
First did auto captioning using florence2 in comfy with an image batching which gives a very detailed caption including clothes, etc. for each image.
Blacked out all the faces in the training data set (about 40-50 images). Literally painted over the faces so they were black. (Since then I found I can use batch image processing in comfy with a segmentation node).
Went back and altered the captions - instead of for example "a woman wearing a white dress", I changed it to "A woman with a blacked out face wearing a white dress" (VSCode is useful for this). This means the blacked out face is a separated concept during training.
Then I trained using fluxgym (I use quite high dim-alpha 64/32). All FP8 in this case. (I alter the scripts manually - much easier that the UI)
Like magic, When you generate images it forgets about the black mask and uses a real face of a described woman.
Only issue then is competing LORAs. You can get a good pose, or a good face, or both average. Hopefully you get a good result at 0.7 pose 1.0 person lora.
Am trying other poses, but none as good as this one so far...
spudajt401 You'll achieve much better results with masked loss training. You can batch process all the training data too - you'd just be creating separate black and white masks and referencing those in training.
272727 Thanks. I didn't know about that. So much interesting stuff to learn here. Not sure I can do it in fluxgym, but I'll definitely try it.
spudajt401 no problem! Thanks for sharing your LoRA. I've never used fluxgym but I believe it's just using Khoya as the back end so, depending on the version that it uses, it should definitely be possible. As you can imagine, being able to train only the elements of training data that you want without other stuff being lumped in with the concept is a gamechanger. So definitely worth looking into. :)
@272727 Does AI-Toolkit support that approach?
@Jellai I had my old dataset from the Flux lora I did with fluxgym a few months ago, so thought I'd try it and it worked fine with Ostris AI Toolkit. I have other loras I have done with proper masks on kohya-ss, but the blacked out images to train z-image seem to work fine.
I know art when I see it.
Work well! Great LoRa!
This looks so ugly and unnatural! Like all other pussy loras for z-image these days. Thanks for the training but completely waste of time... ;-)
Can you read? It's a pose LoRA not a pussy LoRA. So many 'geniuses' lately...
@TheGlowingGuardian Oh, someone's a bit sensitive about the truth. It's not your fault the pussies look awful; that's the model's fault.
And yes, of course I can read! But that doesn't make it any better that he is offering a pose Lora that nobody needs. Z-Image follows prompts so well that you can create any pose instantly. Three or four words are all it takes for this pose. So it's pure download farming... 👎
@MrSmith2025 I'm not OP, 🤣 genius.
@MrSmith2025 dude use inpaint and stop bitching.
@challasvarias01106 What does Inpaint have to do with not needing Pose LoRa for Z-Image because it can already do that natively? Stop using drugs, please... They aren't good for you!
@MrSmith2025 Then your bitching is like this LoRA, useless.
@TheGlowingGuardian What exactly is "bitching"? I'm an adult, not a little girl from your kindergarden playgroup! And now MrGoldenGuardian go and find some pedophile child pics and report them instead of dropping stupid comments...
@MrSmith2025 Fascinating subject. You'd probably learn some social manners and respect getting right back to kindergarten.
These LoRAs are trained on a distilled version of Z-Image, a model not intended for NSFW material. Until the base model is released, we can't fault people for trying.
I'm the OP. Please try V2 (Z-Image only). Hopefully you will get better results. It is trained on higher res images with a new dataset. I think the genitals look better, but its never as good as I want it unfortunately.