Presenting DR34MSC4PE's new WAN LoRA:
DR34M15H - Multi-perspective, hyper-flexible missionary position (POV+)! Should also have no issue handling sticky climaxes either, just prompt accordingly…
Surprisingly, it also handles prone bone (and more), just specify facing away or on stomach, etc. Give 'er a twirl ... ;)
===========================================================================
7/22 - It's here. T2VHDv1 - Enhanced for larger resolution generations, multi-res training, better dataset balancing! Available on early access for the next three days, free to all after that! Enjoy!
7/15 - Thanks for helping BEAT our donation goal. Enjoy the model - more soon!
7/13 - Harder, Wetter, Faster, Longer - Enjoy this summer treat, from us to you. T2V/I2V-480p, coming soon too!
Please remember to like and share - if you like our work, give us a shout (or buy us a coffee) below!
==========================================================================
Sample Prompts:
m15510n4ry, A shot of missionary sex and a detailed view of a woman's vagina and a man's erect penis which is penetrating her. The woman's skin is light and the woman having sex has natural pubic hair above her vagina. The penis is erect as it thrusts in and out of her vagina. There is a slight sheen of moisture on both the penis and the surrounding skin. The background is a fancy bedroom where she is at the edge of the white bed and he is kneeling between her spread legs. The lighting is even highlighting the natural skin tones and textures.
m15510n4ry, A close up shot of missionary sex and a detailed view of a woman's vagina and a man's erect penis which is penetrating her. The man's hand is visible at the top holding the base of his penis. The woman's skin is light and slightly textured with visible folds and natural coloration around her vagina. The penis is erect and below average size as it thrusts in and out of her vagina. There is a slight sheen of moisture on both the penis and the surrounding skin. The background is a fancy bedroom where she is at the edge of the white bed and he is kneeling between her spread legs. The lighting is even highlighting the natural skin tones and textures.s
===========================================================================
DR34MSC4PE is
c0ur4ge (trainer/qa/inference code) /
ERA5ER (trainer/captions/tooling code/data/qa)
Be sure to check out our FLUX model - C4PACITOR!
Like our work? Buy us a coffee: https://ko-fi.com/dr34msc4pe
Please generate responsibly!
Description
We out here - here's the T2V_HD variant of DR34M1SH. Stay tuned for new updates and more variants! =)
FAQ
Comments (45)
Great work! Would you mind my picking your brain on your training for a moment? I see 42,000 steps for 360 epochs, so roughly 117 steps per epoch. I'm curious, if you using a repeat of 1 for your training? Does that mean you had about 116 in your dataset? If so, how many videos vs. images, and most importantly, what FPS was it encoded at?
Sorry for the asks, just impressed with the work.
I should note that the step count is usually a guess - the epoch is generally correct. This dataset was roughly 100 videos. More or less using standard diffusion pipe settings. We’re not really doing anything ground breaking from a hyper parameters perspective- as always, it’s your data and frequent sampling (and the occasional interpolation) that sustains quality.
c0ur4ge Thanks for the additional info, on my original question, you just doing 16 fps though, yeah? My original ask around dataset value and repeats is because your step per epoch is really low. With a dataset of 100, that means you're likely doing repeat of 3 with a GAS of 4, yeah? I'm asking purely because there's a growing split in the train of thought on batch values for wan it seems. Some of us trainers who ascribe to the dataset repeat of 3-5 per batch of 1 (maintaining gas of 1), so they get more steps per epoch, and a growing number of others it seems who are training at batch count of 1 with repeat close to 1, so they land at low step per epoch (like yours appears to be).
KinkMaster We wiggle this a little but mostly it’s the Diffusion Pipe defaults - I believe it’s a BS of 4 per pipe x N GPUs. I think default GAS is 4 for our runs and typically N = 2. :). No repeats but that’s because I’d rather keep the epochs clicking for more granular sampling vs. overtraining within an epoch.
c0ur4ge thanks for confirming, that largely was my guess based on the counts!
KinkMaster No problem - if you think this is apex, just wait till you see what’s next. Early samples have me in the “how the fuck” stages still - lots of sampling to do but yall will have it soon. Stay tuned!
KinkMaster The videos are exported at 29.97, around 2-3 seconds long each.
This might be the best LORA ever. Great job!
We’re glad you’re digging it - more real good stuff coming soon!
Nice one! works really good for Anal as well!
me waiting for this t2v timer ... 4 more hours ...
occasionally blurred out watermarks appear in the bottom right. you might want to prune this kinds of video from your data set.
otherwise fantastic lora!
edit: having so much fun with this, it pairs well with other loras, can't wait to see what else you have cooking
We’ve caught this a time or two in widescreen formats. We’re shuffling things around a bit but ultimately some is a bit of a necessary sacrifice for quality to include some datasets. Good call out!
c0ur4ge I haven't tested your LoRA yet, but I had the same issue with one of the LoRAs I trained, thinking blurring the watermarks would eliminate them. Turns out you would just get blurry watermarks in every generation. What fixed the issue for me was simply captioning them in the dataset rather than blurring them out. (i.e. This video has a watermark in the bottom right corner) And then just put "watermark" in the negative prompts in ComfyUI.
schwadorf Eeeyup, that's exactly it - in fact, we normally always do this but it got missed by the captioner.
One thing I noticed, cum remains static no matter what I prompt. I am assuming this lora has no trigger for it anyhow, but others usually work for that effect.
I've tried multiple on the side to try to achieve movement but could not get the effect. I found only two videos in the user examples that managed it, and I still couldn't get anything to move using similar prompting and loras.
I see a few user examples with similar issues, particularly anime/illustration-style starter-pics.
I haven't yet tried to lower strength on this one though, as it shines on 1. I've been using high CFG as well. Often that produces stop-motion but it's been pretty good on high settings.
I'll keep trying XD
Edit: it's especially hard to try stuff on site since the big Wan-Lora tagging fuckup as well, so that's not helping. There is a lack of usable 720p loras since they messed that up.
I appreciate the feed back - I have a suspicion on this and it has to do with potentially the clips not training to the end of the (already shortened) clips and might be missing some of this "content" - I'm going to ensure the entire clips are trained or trained in splits in subsequent runs but, that all said, we've seen it work in images shared. I'll see if I can reproduce the issue.
c0ur4ge Nice. I'll summarize my point in case anything was lost in my garbled comment: when using i2v with pictures that already have cum/liquids on them, the liquids stay static in the output.
Very flexible lora. Really it is a hip thrusting lora, and you could put the penis anywhere. good for mixing at low power with BJ loras to get the mans hips in the frame and thrusting.
can use with no dick in image?
Also want to know this
yes.
I am not able to run this due to metadata PG13 restrictions. Anyone had luck run this on your local computer? whats the hardware requirement?
if you download the model an run on your computer, will that work?
its primarily for that lol
I wonder how to create the first image that looks very realistic and has a right pose. Thank you.
illustrious, pony , flux dev with loras… then wan I2v
i gotta keep playing and learning with this one
This LoRa is godsent, thank you. Due to it's versaitility, you save yourself a lot of other LoRa's when using this. It also seems to do fine with the Lightx2v LoRa and Wan 2.2.
One critique point though: I personally don't like LoRas with fancy 1337-Trigger words. But that's a sacrifice I am willing to make ;)
One thing, sometimes the penis slips out of the vagina and keeps thrusting the air. How to prevent that? (Using the I2V)
What’s your negative?
@c0ur4ge all fine, was my fault, for testing with Wan 2.2 I increased LoRa strength to 2.0 and forgot it...
After turning back to 1.5 all was fine
@haenlesn937 even that’s pretty high lol
@c0ur4ge it is, but for wan 2.2 i somehow need to turn this strength up or else i just get poking instead of thrusting. Mayber that's also the SpeedLora's fault, still testing around
@haenlesn937 Try to disable the speed lora on the high model and only have it on the low model. Seemed to work for me in T2V at least.
oh, this model just disappeared from i2v. It worked so well with i2v, why did it disappear? That's sad.
Must have been because they nuked all my images (I'd used a couple actual internet images which turned out to be real/copyrighted but they deleted all of them erroneously). Just re-added a single one and re-published it.
@c0ur4ge This model is really awesome! Because it has the sharpest rendering of genitals in i2v of all models. Reverse Cowgirl is good too, but Multiperspective is mush more sharp much more often. The best video model on civit!
I've yet to run a WAN2.2 workflow that wasn't rife with issues. Sometimes I'll get very good results, but this "high" and "low" stuff with no clear indication of what strengths work best for all of these models has made image generation a frustrating experience. WAN2.1 seems to actually be better for consistency.
I just tested WAN2.2 vs WAN2.1 using the same imagine and prompts and WAN2.1 nailed the prompt each time whereas WAN2.2 comes out distorted, low quality, or high quality but doesn't adhere to the prompt, and so on. Until I play around with the weights for an hour or so I simply get garbage outputs.
And the premade workflows from other people always consist of some custom or obscure node you can't get via comfyUI manager or if you can it doesn't work properly half the time.
It'd be nice to have a workflow for WAN2.2 that isn't setup to run on a 5090 and doesn't use 100 nodes that you need to download from outside of comfy manager.
Any chance we're getting WAN 2.2 standalone version of this?
Yup! Although the forthcoming DR34ML4Y may sate your need since its got all of our concepts in one - it's just the slowest to train. Nearly finished and then I'll round out the single concepts. :)
Probably still coming on this - just a couple in the hopper.
I dont understand how you can get so improuved mouvement, i get only few mouvement