About this version (qwen v2)A brushup of the dataset, with better result than V1 for qwen. I will use the same dataset again for zturbo when the base model is released. Trigger word: PENISLORAWhat can this lora do?This lora can add erect penises to both men or women viewed from the front/side. Other angles such as POV may have a backwards penis head. Other things it can now do: Side view of the penisCumming / CumshotsBlowjobs (its captioned for the words "blowjob" and "deepthroat" )What can't it do?No penetration in the training data. Also nothing from POV angle, though there is a few images from above and 1 POV video in the training data.Sometimes blowjobs with cumming have the penis slip out the closed mouth.Recommended SettingsIt works pretty good with the new lightning dyno high model. I'll link to it in my example workflow. I like to use dyno high model (no lightning lora), then for low I use the lightning v2 lora on the regular 2.2 low base model.Dataset84 images at 512x resolution43 videos at 256x resolution(I let DP pick the aspect ratio automatically)This is the same exact dataset as the 2.2 5B model. I made no changes.TrainingI used the default diffusion pipe settings.[optimizer]type = 'adamw_optimi'lr = 2e-5betas = [0.9, 0.99]weight_decay = 0.01eps = 1e-8I was baffled why it was taking so long to train the high until I realized after over 60 hours of training that I had put my videos in the images directory which resulted in the high being trained ONLY only on videos and twice (once with a very high resolution). Once I fixed this, I went back and trained from 11K steps up to around 13K with the images in the training data. The high model was fine without to be honest.For the low, I trained it properly with videos and images the whole way, around 6K steps in I upped the image resolution from 512 to 1024 actually and didn't get an OOM (it fit around 24GB exactly). I trained it to around 10.5K steps. Also I trained the low on the full timestep range (0 to 1 instead of 0 to 0.85) from some advice, it may switch better over from high to low on the speed up lora with low steps.I think I might do another version with more angles such as POV and from the behind to make this work for any situation. In that case I don't think it needs 10K steps per training session, epochs around 5K steps looked fine.The resultsI think it was a combination of improved captioning and 2.2 base model being better. But this lora turned out really well.

Penis Lora (+Blowjob, +Cumshot) [Taz]  - WAN 2.2 14b / 5B / 1.3b T2V & I2V (Wan 2.1 & 2.2) + Qwen + Zimage Turbo

Update memo:This update took almost 20 hours of training on a H100S GPU on Runpod and cost me around 50 USD so I appreciate any feedback and buzz :). I do this for learning and passion and I have decided that as long the training isn't regressive, I will try to make this into a one-stop solution for erect penis in Wan t2v 14b. (if I have enough funds to keep on runpod or I decide to locally do it on my 3090...) That said this time seems much bigger improvement give it a try, you will be surprised compared to v.05 and with such little data added. That said this still BETA and not perfect...Check the update notes below for details of whats new.This is a lora trained on Wan 2.1 14B t2v model and is designed to give frontal shots of average to large size erect penises for women. (I have not done any generations with men).Important info:Trigger word: PENISLORA (you can just chuck it at the front of the prompt)Best Strengths: 0.7-0.85 (anything higher you get less variety, lower less penis). I prefer 0.85.Version update v0.6 notes:<ul><li>The data I've now added contains videos/photos of blowjobs, penis stroking, and a few more angles. </li><li>This is an additional 6K steps / 60 epochs on top of the 3.3k steps / 42 epochs from v0.5. </li><li>You can now show the penis from the front from most angles clearly (less weird tips but still not perfect). </li><li>It has 1 POV BJ content but its still not enough, cannot do POV really yet, but can do from the side or front.</li><li>I think now it shines in terms of gripping the penis or putting it into the mouth.</li><li>It can deepthroat, and should mention "tip of penis" for the head and "base of penis" for the base.</li><li>It still has some issues, you can get bent or stretched dicks since it really wants a gap between the top of the hand and the tip of the penis.</li></ul>Technical notes about v0.6 training:<ul><li>Loaded LoRA weights from Epoch 42 via init_from_existing (I will do resume from checkpoint next time but data seems fine).</li><li>Reset global step count to 0, starting a new training run with inherited knowledge</li><li>Added 8 new captioned videos + 8 images into the dataset (same resolution)</li><li>Maintained prior dataset entries for continuity and reinforcement</li><li>Did not resume from full Deepspeed checkpoint (global_step) — used LoRA state only</li><li>Continuing training from loaded LoRA does not poison data, safe for extending behavior and loss is still good range.</li></ul>Workflow:I also will include a workflow .json file so you can see what nodes and settings I have. For the most part its just Kijai's standard wan 14B t2v fp8 model optimized for use with 3090 (tea-cache, slg, torch, sageattn, CFG star zero, etc...). For the sake of genning as much as possible I set my frames to 65 and use the wan default 832x480 resolution with interpol frames / upscale I get maybe around 10 mins or so per 4 sec clip.Sample prompts:A high quality video of a nude woman sitting comfortably on a wooden chair in a softly lit bedroom, PENISLORA fully erect and resting against her belly as she gently strokes it with one hand, the camera fixed directly in front of her at chest level, subtle body movement as she shifts slightly in the seat, warm and intimate atmosphere, realistic styleA high quality video of a beautiful Japanese woman with large natural breasts lying on her back on a bed with soft white sheets. She has fair skin and short deep blue hair in an ear-length bob with a hime-style cut, featuring straight bangs and cheek-length sidelocks. She wears a black-and-white maid outfit with the top unbuttoned and pulled open, exposing her bare chest. Her erect PENISLORA penis is fully visible, attached naturally at the base of her pelvis between her thighs. She smiles warmly and looks directly into the camera, her expression relaxed and confident. With one hand, she wraps her fingers firmly around the shaft and moves it slowly up and down in a stroking motion. Her balls rest softly against the bed. The camera is positioned directly above her, capturing her full body from head to pelvis, with her entire face clearly visible in the frame.A high quality video from the point of view of a passenger sitting on a public city bus, directly facing a beautiful Singaporean woman seated across the aisle. She wears a tight white knit turtleneck sweater with the subtle outline of her nipples clearly visible through the fabric. A short pleated skirt rests on her thighs, and her legs are slightly spread. As the bus bumps gently over the road, she opens her thighs wider, revealing her erect PENISLORA penis, naturally attached between her legs and swaying slightly with the motion of the vehicle. The shaft is flushed and firm, the tip fully exposed beneath the lifted fabric of her skirt. She locks eyes with the viewer and smiles intensely, her expression playful and focused. Around her, other passengers are present, facing forward or looking out the windows, visibly ignoring the scene. The interior of the bus is warmly lit by late afternoon sunlight, and the camera remains fixed in a frontal seated perspective.A high quality video of a beautiful Japanese woman with large natural breasts standing behind a coffee counter in a cozy modern café. She has fair skin and short deep blue hair in an ear-length bob with a hime-style cut, featuring straight bangs and cheek-length sidelocks. She wears only a dark green barista apron tied around her neck and waist, leaving her chest fully bare. Her large natural breasts are soft and fully visible above the apron’s low neckline. Her erect PENISLORA penis is attached naturally between her legs, with the tip peeking out just below the hem of the apron. She stands with her legs slightly apart, smiling warmly and looking directly at the camera. With both hands, she holds up two takeout coffee cups, presenting them proudly. Her balls rest gently behind the fabric of the apron, visible at certain angles. The camera is positioned above and angled slightly forward, capturing her full body from head to knees in a bright, warmly lit Starbucks-style setting.Negative prompt:色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，倒着走, 模糊的眼睛,低细节面部,水印, 图层, 故障, 面, 屏幕残影, 图像干扰,Background:<ul><li>This was trained on 28 images at 512x512 and 3 videos at 640 x 480 (16 fps ~65 frames)</li><li>I trained this using the excellent tutorial and runpod created by <a target="_blank" rel="ugc" href="https://civitai.com/articles/12330/wansdxlflux-all-in-one-lora-training-diffusion-pipe-with-auto-captioning">Hearmeman over here</a> Settings is the default settings of his, except a few things like I set repeats to 5 for the images. Also shoutout to <a target="_blank" rel="ugc" href="https://civitai.com/models/1404755/studio-ghibli-style-wan21-t2v-14b?modelVersionId=1587891">Seruva19's Ghibli lora</a> and his detail and shared training data, without that reference I wouldn't have done this.</li><li>Video setting buckets I added a custom 61 frame setting among the 65 already there.</li><li>Images and videos I cropped the faces out for the most part though 1-2 images do have faces.</li><li>Ended on Epoch46 at around 3.5k steps. I didn't see any issues with loss and I feel like this could be trained for another few hours without issue.</li><li>This took around 10 hours of training on an H100S on runpod and cost me around 33 USD, so appreciate any buzz :)</li><li>Also the videos are all cliped and converted to 16 fps and around 3-4 seconds each. This was a good learning experience for me.</li><li>Captions are one run through joy caption for the images with manual touch-ups by hand for wording consistency (gendered all to female pronouns, changed semen to cum etc.). Videos were captioned by hand.</li></ul>I made this because while the erect penis loras already here for Wan are great, I didn't like some of things such as gigantic dicks only, surgery scars, trained on men, or aggressive tan lines. This data set is only transgendered women and I tried to get photos of women with more natural and feminine bodies/breasts and not overly huge dicks (though there is some in there). The videos also feature cum shots and masturbating but the LORA in its current state isn't quite good at either. I would recommend this for more for jerk free or non hardcore videos. Though I've had some success with both jerking and cum in the videos. This model really shines on close-ups and lower camera angles going up. That said, this LORA works perfectly fine with men so don't worry.Some tips (updated for v0.6)<ol><li>You need to anchor the penis to the body in the right location. For example include: The woman’s erect penis is naturally attached between her legs.</li><li>Cum is now trained a bit better, it will go out of the tip of the penis, and is more cum-like than last version. Still though it needs more training. Their mouths remain closed and dry, with no fluid near their lips. Thick white cum shoots from the tip of her penis,</li><li>The lora now has various angles like profile and above but still not able to do POV angle and the penis head is biased for frontal view. I will try to get more out of it in future versions. You will get some wonky looking dicks depending on if they stray too far off from the front.</li><li>I tested it with other loras and for the most part pretty flexible and works fine. But you may wanna play around with the strength when mixing.</li><li>This new version tends to elongate dicks, but you can get smaller dicks. And you may have to play around with prompt or seed to get not overly long dicks sometimes.</li><li>Mouths now tend to naturally move towards the penis and suck on it. It is also trained and captioned with "deepthroat" words and content.</li></ol>Warning:I labeled this as a "BETA" because I feel its unfinished, you can get some real great realistic looking dicks out of this, but its very easy to get deformed monster cocks too... I will try to fine-tune and also just pick up training again. My next big project is to tackle something like the Ghibli lora for another anime I like, this lora really helped me learn about how to prepare data set and caption videos as noob to WAN training. Please provide constructive feedback if you like, and I can work on getting this lora in a more stable and useful state, I think though it can get the job done in its current form so please give it a try.

Files with SHA256: e21caef826384e2b979e84fa0e01a861c470e5cc3589c18ed40cf101a5f1742b