⚠️NOTE! Many of the preview videos use other loras as well (fingering, missionary, cumshot). This is to show how it stacks with other loras. This lora is only toe-sucking / foot worship. See metadata on previews for other loras used!
This lora is kind of a brat. Don't expect perfection...
Weight 1.0~1.3
1.0 is usually fine, but you can experiment with higher weight for better results... sometimes.
Works for self or partnered toe sucking, gender-neutral.
Works best with start images containing feet and face already in close proximity. There are plenty of foot worship image loras to help make start images.
Works best with photorealistic start images. Anime style does work, but the toes may not move as much, and sucking action may look weird.
Also works with start images without feet. But the feet that appear may look weird. If you try it with a footless start image, you need to prompt something like "she reaches down and lifts her leg and foot to her face."
This lora is sensitive to parameters- play with your CFG, shift, and resolution. I had best results with ~0.8MP resolutions.
Description
FAQ
Comments (15)
Aaww maaaan you should have included the whole thing. Toe sucking, sucking between toes, sole lick, smelling soles 😭🙏😭
It will do most of these things. Maybe not sniffing... the training data included very diverse foot worship clips.
Give it a try, licking soles is probably the easiest thing to accomplish with this lora, just prompt for it.
I focused on toe sucking for the previews because it was the hardest thing to achieve.
HOLY! TAKE MY BUZZZ!!
T2V pleeeease!
With how complex these actions are, and how amazing the result was, I feel like this set a new standard for loras. Very interested in how you managed to train it so well.
thanks! though tbh I'm personally not totally satisfied with the result 😂
I'm certainly no 'expert lora trainer,' just a weirdo doing this as a hobby after work. But since you asked... training was not much different than other loras, just put more effort than usual in curating the dataset and ended up training way longer than usual.
I had a ton of clips but settled on 81 general foot worship clips, with mirrors = 162, 16 of those had "perfect" motion so I set those to repeats=3, so effectively 194 clips for high noise.
Low noise had 73 clips which were all close-ups of licking, sucking, etc- lots of toe and tongue motion.
I actually cut down and cleaned up the dataset significantly from my first attempts which were 300-something clips. I still may have had too many, but the motions were similar enough I figured (hoped) it would be okay.
High noise was trained to 12,804 steps and low noise to 5,986. Which is way more than I normally do, but testing every ~1,000 steps the results were actually getting better rather than breaking down. I probably could have kept pushing it even more. iirc loss settled around 0.04 but I didn't keep track unfortunately.
Both models were trained with 640x360 res clips, higher res than I would normally do for a full-body motion lora at least for the high noise model. All clips were 3 seconds.
Training was with musubi-tuner on a 5090
@icelouse Oh damn. 640x360. Must have done a lot of block swapping, even on a 5090. How many hours did this take total?
@Jellai lol yes, swapped 32/40 blocks. I didn't keep track of the total time for the final published version, but estimate maybe 24 hours for high noise and 12 for low. I went thru several iterations... train one night, test for a couple days, resume training, test again, get pissed, nuke it, trim dataset and start over... basically the entire month of September training and testing this one so I stopped keeping track 😆
@icelouse what, this big resolution, the f***! How long did it take to train? Thank you so much for the LORA, appreciate a lot! EDIT: you already answered the time taken, this is ... really long. I did some LORA for SDXL but the most time I spent was curating images.
works very well !🙏
i used the prompt you gave and it makes my character eat a chewing gum with her mouth open after sucking for 1 sec lol any fix ?
Will there be a version for Wan 2.1?
Hi icelouse, anyway I could get your workflow for this lora? My results are not coming out great. Thanks!
Anyway we could get an LTX2.3 version?
maybe 😅 I just started playing with LTXV23 this week. Working on my first lora now, but I might try making this one next
Details
Files
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
Mirrors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
w22_i2v_FOOT-WORSHIP-TOE-SUCKING-v1_H.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
舔足H.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
舔足H.safetensors
wan2.2_i2v_highnoise_FOOT_WORSHIP_TOE_SUCKING_v1.0.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.