v2.0
Version 2 adds 150% more training data and more motion overall. This is now a general purpose POV implied sex LoRA, not strictly limited to "riding from below."
Depicts a woman bouncing up and down, with optional breast play. Key phrases:
From a low angle, a woman is having sex, bouncing up and down.From a low angle, a woman is having sex, bouncing up and down. A hand reaches up from the bottom and grabs her breast.From a low angle, a woman is having sex, bouncing up and down. Fast and dynamic motions.
{she leans down hard until her face hovers inches from the camera, straddling close enough that her hair falls forward and brushes the lens as she pins her weight downward|she throws her hair back with a sharp motion, then drops her chest forward toward the camera in a controlled press, keeping the straddle tight and dominant|she plants both hands beside the camera and shifts her hips forward in a firm pin, lowering her face until it fills the entire frame|she hooks her hands behind her head and bends at the waist, pushing down into the viewer’s POV with a confident smirk as she maintains the mounted position|she drives her weight forward from the straddle and reaches one hand toward the lens as if steadying herself mid-pin, breath warm against the glass|she lowers herself abruptly, bringing her chin nearly to the camera while her knees grip tighter around the POV, hair swinging forward in a quick arc|she braces one arm beside the camera and leans in until her forehead almost touches the lens, applying pressure as if keeping her opponent down|she rolls her shoulders forward, collapsing her upper body toward the POV while maintaining the straddle, eyes locked in an unbroken stare just inches away|she shifts aggressively forward, brushing a strand of hair aside as her face crowds the frame and her weight presses down with intent|she thrusts both arms toward the camera as if adjusting her balance during a pin, her face dropping close enough that only her eyes and mouth dominate the view|she dips suddenly into a low lean, winking from inches away while her body stays tightly pressed over the POV, knees firm around the position|she extends one arm past the camera and pulls herself even closer, lowering her face until the POV feels fully pinned beneath her}Description
FAQ
Comments (20)
Really good!
Can you please share some details about how you trained it, how many videos and other details?
Keep up the good work! ❤️
It was trained with 10 videos, hand-captioned using natural language. I used AI-Toolkit to train locally, using all default settings for WAN 2.2 14B I2V. All clips were 80 frames. I let it run for 3,000 steps. That's about it, but I'm happy to answer any other questions.
@_Species8472_ Thanks for sharing! this is VERY impressive consider only 10 videos.
Did all videos uses the same FPS and RESOLUTION? can you please expand on that?
1 - In example: all video were 30 fps? or some videos used different FPS?
2 - Did all videos used same Resolution? and what resolution did you use?
I'm asking for trying training my second test for Wan 2.2 LoRA
because I try followed Ostris (AI-Toolkit developer) YouTube tutorials and my results were BAD even on 5000 samples... and you're saying you just used the defaults.
If you can expand on any extra information it will be very helpful, thanks ahead!🙏
Also please keep up the good work, your LoRA is really good!
Much Love to You!❤️
@Virtual Here's the dataset info:
4.mp4: 173 frames at 23.98 FPS (1920x1080)
8.mp4: 155 frames at 23.98 FPS (1920x1080)
2.mp4: 145 frames at 23.98 FPS (1920x1080)
0.mp4: 143 frames at 23.98 FPS (1920x1080)
3.mp4: 126 frames at 23.98 FPS (1920x1080)
7.mp4: 123 frames at 23.98 FPS (1920x1080)
1.mp4: 122 frames at 23.98 FPS (1920x1080)
9.mp4: 111 frames at 23.98 FPS (1920x1080)
5.mp4: 110 frames at 23.98 FPS (1920x1080)
6.mp4: 110 frames at 23.98 FPS (1920x1080)
Everything was trained at 256 resolution. I'm really liking AI-Toolkit now (I use the UI). I was using diffusion-pipe for a long time, but I could never get it dialed in for WAN video LoRAs.
@_Species8472_ Hi, ty for info .. how long did it take and what is ur specs ?
@Useful_Ad_52233 8 hours maybe. I'm on a 4090 with 48GB vRAM.
@_Species8472_ Thanks for sharing! now it make sense why it's still slow on 24 and 30 fps because it did the average speed I guess based on the amount of images from 100+ to 81
Did your videos exposed FACES with expression?
Because it does a GREAT job not making similarity issues and also they are NOT talking like in other LoRas... I wonder if on your DATASET videos their faces are talking? or if you CROP their faces?
On AI-Toolkit:
Did you CHECK ALL resolution options? or only 1024 for example?
@Virtual I just uploaded the training data so you can check it out. I only trained at 256 resolution.
@Virtual One thing I notice from the training set is that there was a single video that I forgot to fill in the caption for, so that one was captioned with a blank text file. That was entirely by accident.
@_Species8472_ Thank you so much for your kind tips & tricks! I can't wait for your upcoming LoRA and also learn from your training experience will be great!❤️
@_Species8472_ So if I understood from this specific LoRA training at least:
- Dataset: 10 Videos - 1920x1080 - 23.98 FPS
- Resolution check ONLY 256 Resolution
- Ignore caption of the face, right?
Since I'm about to try train in AI-Toolkit, there are resolution to check and by default 256x256, 512x512, 768x768, etc.. so for MOTION / ACTION it's fine to train on 256x256, your LoRA proves it 👍
Am I missing something? any Tips & Tricks will be appreciated 🙏
@Virtual That's all correct. The only other thing is I set the frames to 81 in the AI-Toolkit config.
@_Species8472_ thank you for all the sound advice, i'm also trying to start training my own wan 2.2 loras. wanted to know how exactly did you format the videos to your liking? (probably a very beginner question but I'm just trying to get everything ready) Like did you use premiere pro or any other video editing program?
hmm..., if you get leveled image, which is not shot from low angle u get POV standing sex...just sayin'
Trained very well, a lot of actions, looking forward to more poses.
This prompt layout where you organise prompts with { and |, how do you use that? With some specific node? What is the workflow there exactly?
Those are dynamic prompts. You put them in the "Random Prompts" node from the https://github.com/adieyal/comfyui-dynamicprompts extension. You can download any of the sample videos to inspect the workflow.
Amazing, one of the best that I tried. Please we need the same for back/reverse position, I tried with this and it doesn't work, turn the face like front
dude we got both angles?? you're the best.
Bouncy fun!
Details
Files
povfrombelow_WAN22_copy_3000_high_noise.safetensors
Mirrors
high.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
POVRiding_HIGH.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow-high-i2v-wan22.safetensors
POV_Riding_from_Below__i2v__HIGH_.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow_WAN22_high_noise.safetensors
wan22_riding_povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow_WAN22_copy_3000_high_noise.safetensors
povfrombelow_WAN22-HN.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.