DR34ML4Y by DR34MSC4PE
This LoRA, incredibly, manages to handle ALL of our specialty LoRA concepts - Double/Single Blowjobs, Missionary, Reverse (and regular) Cowgirl, Doggy, wrapped in one model! It obeys very well and is largely extremely accurate to the given keyword/position! Don't be shy; give it a try! =)
===========================================================================
Update - Due to civit making sure we do not get paid for the use of our model via generation any longer, on-site generation has been switched to 100 buzz per generation. Please voice any concern directly to Civit AI's support over their change effective 11/1.
10/15 - H3 PREVIEW: we're so back.
Note: this is an early preview - expect some T2V challenges. H3 is not WAN/LTX. Prompt accordingly.
Turbo LoRa moves latent space away from NSFW concepts. Results, especially audio, will get worse the more turbo you combine into the mix.
AS ALWAYS: Enjoy, generate responsibly and share what prompting tricks you learn along the way!
==========================================================================
Note - Missionary and cowgirl can have some mix-up if you don't specify the woman is on her back in the prompt.
===========================================================================
DR34MSC4PE is
c0ur4ge (trainer/qa/inference code) /
ERA5ER (trainer/captions/tooling code/data/qa)
Please generate responsibly!
Description
First version of our AIO for WAN 2.2 TI2V - Enjoy!
FAQ
Comments (25)
waiting for i2v 14b version
That's next!
c0ur4ge NEED IT IN MY VEINS
ablestable69420 It’ll be in the oven by this evening. Likely next weekend by the time it’s finished and I’ve selected a suitable candidate. Keep yall posted. For now, try the 5B - it’s much higher quality than I expected and FAST. Not perfect, but damn good imo.
Wait no longer - it has arrived :)
5B? are you cooking 2.2 wan low + high too?
Yup - next in line. Life stuff has kept me busy so haven’t been able to kick this one out the door until today (and I don’t like to be trying to train/eval > release at a time).
Would be awesome to see a 14b high/low
SO glad you're doing 2.2. Very excited to see what you can do with High and Low.
how do you get a better quality penis? or do you have an additional lora? they just luke like flesh tubes, not very real.
pretty sure you're just stuck with the one you're born with, man
but seriously, it probably is a settings thing with your sampler if I had to guess if you're consistently having issues.
c0ur4ge yeah im just getting body horror. very high ddetail body horror but its not missionary by any stretch. Running wan2.2 i2v 14B+res4lyf clownsharkksampler high and low noise passes.
@jwcoop3r631 there are some penis specific lora to add sometime it helps. Many lora give very good results when the starting image o prompt are how expected but deform the penis when not, sometime even simply by the guy entering frame from left instead of right.
In that case adding a penis specific lora help just some try with different weight is needed to preserve both the action and the "penis restructuring". Never tried but probably help using different weight in low and high. Example using the penis lora mostly on the low noise side.
I'm lazy and my workflow is fast to prompt but do not have the possibility to use different weight for high and low :D
@jwcoop3r631 I love sharksampler but it might be that.
Hello and thank you for the work.
Wan2.2 it2v 5B: I can't get decent images.
This 5B model is a horror (on my M2 Mac). the faces are catastrophic.. (I have no problem with the other models..) if anyone has a solution. My images are rotten, even without lora, with the 5B THANKS
I'd say that I'm not sure it's really as useful as a purely image model as the other WAN versions (not saying it can't be without some workflow help) but to clarify: Are you talking about your videos? Or images specifically?
c0ur4ge Thanks for your reply. I'm talking about videos. I can't find a correct setting. Following my message, I tried to produce in 1280x704 size. Strangely, the vram is well managed and I can make 4-second videos. The quality improves significantly, but for the nsfw, when I use a lora, it breaks everything. I tried changing the weights, from 0.10 to 1.. I use euler simple. I'm on Mac M2 pro 32GB. This head is translated from French to English in Google Translate. My English is not good :-)
charlie_perso C’est d’accord! I’d ask - are you using the default WAN workflow? I’d probably literally start with ComfyUI’s WAN image2video workflow and add the Lora node after the model loader and see if you see the same.
I expect this is because of a speed Lora or node that is trying to cut corners for you.
c0ur4ge I used the default 5B workflow from WAN. and I added the Lora. I have no problem with the 14B models. Before answering you, I loaded the original workflow. I changed the model and the clip to use the GGUF versions (with the appropriate loaders). Then I added the loraLoaderModelOnly and the lora DR34ML4Y. But the image is still bad
Hi, Great work ;) Do you have an advice for scheduler,sampler,cfg and steps?
it depends on your workflow, i've been learning with rapid model or lora but I use sde 2 gpu with beta or normal, eurler with beta, lcm beta or normal with 4-8 steps, and 1cfg.
without rapid I think euler with beta or simple, cfg 3.5-9, and 10-25 steps depending on if you're testing or not.
Details
Files
DR34ML4Y_TI2V_5B_V1.safetensors
Mirrors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
lwDR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
dr34ml4y_ti2v_5b_v1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors
dr34ml4y_ti2v_5b_v1.safetensors
DR34ML4Y_TI2V_5B_V1.safetensors