IF YOU ARE USING THE NEW BETA GIVE ME FEEDBACK, YOU HEATHANS!
Seriously though, lemme know what's working and what isn't! Also, get a little creative, there is new stuff in the training data. ;) Beta 2 is coming soon, and it's going to have EVEN MORE surprises! Also, send me your photos! IDC if you don't post them here, just send em to me so I can see how it's working with different workflows!
INTRODUCING SEND NUDES V4.1
Welcome to SEND NUDES 4.1, with better prompt adherence, better body shapes, higher resolution training data, and finally, better genitalia.
When I first started working with Flux, my only real goal was to fix the tragic meat button nipples that Flux was infamous for. I am a tattoo artist, and use AI for rapid model prototyping. It's so much faster to get a pose by typing in what I want rather than posing a model in an app. I had created a LoRA called "JUST THE TITS" to remedy the issue, but it soon started to bother me that the output still had Barbie doll parts and that cleft chin. Enter "SEND NUDES" my attempt to rectify those issues. For some reason, female genitalia was a lot harder to train than one would expect, but I think I've got it this time.
SEND NUDES is NOT trained on XXX and will NOT give you all that "Pony Penetration" a lot of you are looking for, but it WILL cut some nice figures for you.
SEND NUDES PRO is a new evolution of this project. The plan is creating the concise dataset needed to properly finetune QWEN into a much better anatomical checkpoint. As the LoRA progresses and my training data becomes more complete, I get closer and closer to being ready for a full finetune. Your tips help me create some of the best and cleanest tunes, thank you for your help!
THANK YOU to all of the BUZZ supporters and LoRA testers that have helped me work on this project, I could not have gotten this far without you!
[QWEN] This lora works well from 0.2 to 1.0 depending on what you need.
[FLUX] This LoRA works well from strength 0.6 to 1.0, but I find 0.8 to give me the most reliable results. YMMV
You can find "REALWOMEN" here"
https://civarchive.com/models/826002/realwomen
You can find "JUST THE TITS" here:
https://civarchive.com/models/708319
Thank you for all the BUZZ and support!
Description
Send Nudes for Qwen
FAQ
Comments (11)
How difficult was working with Qwen? Easier or harder than Flux?
Much less finicky with training data. I still have some learning to do, but it seems like Qwen is a lot more flexible. However, it REALLY REALLY needs concise prompting. Generic prompting will give you generic output. Gotta get hella descriptive.
@ColdNorthMenace You might need to be a bit wordier with Qwen, but I find that it listens better when I give it direction. To this point, Flux has a context length of 512 tokens to fit your description. Qwen has a theoretical max context length of 128,000 tokens (Although reports indicate that it doesn't follow direction quite as well with long prompts).
Overall, I think Qwen is the winner.
@wyldhunt I do believe QWEN is a better model, but the prompts start to fall apart drastically as they get longer, ESPECIALLY with LoRA's. Using an abliterated clip seems to help, I feel like it has a TON of power that the clip is restricting.
@ColdNorthMenace Abliterated CLIP?? How do you use that?
@Dead_Internet_Theory It's a different T5 Clip that is adjusted to allow more freedom in the text prompt.
Can it do male anatomy? I'm gay so I'm not into women but these are high quality. So I could imagine what it could do with male Anatomy. If not does anyone have a suggestion for Qwen that does really impressive with Male Anatomy?
There is no male training here, and I have never done it before, but I suppose I could take a look at it.
Try this out. It's pretty finicky atm, as I just kinda tossed it together. (I don't really have much in the way of dicks in my training images.)
https://civitai.com/models/1898931/send-dudes
Does this work with the GGUF Distill version of Qwen Image?
It should, but I don't know what it will effect.