THE REFINER UPSCALER VERSION IS NOT INTENDED FOR DIRECT TXT-2-IMG USE -- IF YOU USE IT HERE YOU WILL BE DISAPPOINTED.
I've left it enabled if people want to experiment with combining it with other NSFW loras, but don't complain that it doesn't work if you didn't read this. It was trained to upscale only.
DIFFERENT VERSIONS - click on version names above to view:
FLUX NSFW Refiner Upscaler - Us with a img2img workflow (not here...)
!!!!NOT INTENDED FOR DIRECT TXT2IMG USE!!!!! - The closest thing to a 'NSFW Unlock' I'm going to be willing to make. It took a good while to get working. It's not great for direct use, It's meant to be used with the workflow.
workflow here: https://civarchive.com/models/968235?modelVersionId=1084090
HANDSFREEBJ - baked scene, one keyword: handsfreebj
3.5 guidance/20-35 steps/832x1216 size/deis sampler/simple schedulerWOMAN-FRONT-NSFW
WOMAN-BACK-NSFW
TGIRL FRONT NSFW
Training flux for nsfw is tricky, for small scale training, it was having trouble distinguishing between front and back, so I separated nude images of women from the front and back into separate loras and got some great results.
I also did a set of futa images as well (best from the front).
I've not had good luck mixing them into the same image (for the same reason I trained them separately in the first place I'm guessing...)
Datasets were 600-1000 images, training was on a runpod H100 for over 10k steps on each.
Description
FAQ
Comments (16)
cool image
Nice model. For the training did you create prompts for each image in the dataset or was it the same general prompt for all of them?
I used florence to make a mixed tag and caption for each video and had a bunch of custom substitutions for specific words to keep things consistent. I trained on a runpod H100 and i had text encoder/clip training enabled
i get this message when loading:
[LORA] LoRA version mismatch for KModel:
and it just wont load,,,, no errors or anything.////
I'm using it with straight flux. It's working right now for civitai gens as well as comfyui workflows on my 3090 at home. Try a different workflow or to use the base flux model (flux dev 8 bit gguf is what I usually use).
@CautionMergeAhead im using FORGE UI with flux dev fp8 that has all TE integrated..... might htat be?
@liquidmind2013465 try usin " automatic fp16 lora "
@protokAI_ is that a LORA i need to download or an OPTION in FORGE UI?
@protokAI_ ahh i just say the option on FORGE, i will try again then.....
@protokAI_ still "[LORA] LoRA version mismatch for KModel: C:\TEMP\FORGEFLUX\webui\models\Lora\nude-woman-back-v1.0.safetensors"
it only works on COMFY UI, and i dont use it
@CautionMergeAhead it works on comfy UI but not on forge, even with GGUF.... but GGUF will work on comfy.... rare..... still, control net is NOT available on FORGE for FLUX, but control net does works on COMFY, maybe its just FORGE itself.....
@liquidmind2013465 Try v2 that I just uploaded? I used ai-toolkit, not sure about the compatibility issues. Could be solved by using some sort of merge tool or something?
@CautionMergeAhead @CautionMergeAhead the v2 FRONT SHOT works well!!!! i just need the BACK SHOTS now, can you do a v2 version? it is working on forge and conmfy!!!!
@CautionMergeAhead is there any update to do REAR SHOTS V2 ?
@liquidmind2013465 It was working pretty well for me already. If there's a compatibility issue still with forge (which I've never used), you can try merging the lora (averaged) with itself in a1111 or kohya. That will 'rebuild' the lora and hopefully fix your compatibility issue. Unfortunately I don't have time for further fine tuning or troubleshooting atm.
@CautionMergeAhead ok thanks and so kind for replying.... i have used it con comfy and it works great, so good job, regardless of FORGE issues... :D
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.