More LTX foot experiments.
Trained to improve dorsum (top of feet), soles, toe movements, as well as foot/shoe interaction.
It works alright, but don't expect miracles
LTX is just bad at feet, and LoRA can only improve so much when priors are this bad...
Common problems are extra toes and missing toes.
Movement and transitions from dorsum to sole views can get wonky.
Weight 1.0 seems fine, but I've used it as low as 0.3 depending on context and other loras.
I also usually use this with the ltx reasoning lora.
All previews were generated with the official LTX 2-stage Distilled workflow (dev model + distilled lora at 0.5).
T2V previews are mostly marked as such but should be easy to identify.
technobullshit:
I originally only trained for dipping/shoeplay using the same dataset I used for my Wan Dipping lora, except the captions were dramatically expanded using Qwen3VL video captioning.
That turned out pretty well, and also improved general foot knowledge, but it still had zero knowledge of soles, so I added a lot of sole-focused clips and re-trained.
Training data was 102 dipping/shoeplay clips, plus 52 soles/toe-wiggles clips.
Trained at 0.25MP / nominal 512x512 res, rank 32 to 40,000 steps using official LTX trainer on 5090.
The "dipping only" version seems to work better if you really only want dipping, but I'm not publishing it right now because I'm lazy. Leave a comment if anyone is interested and I'll publish it.
Description
FAQ
Comments (14)
Yes, unfortunately, it seems that the LTX 2.3 did very little training on toes.Toes frequently deform into fingers.That's why this Lora is so great.
Yeah, it was so odd when I started noticing that. If someone's foot starts out of frame and ends up being seen later is some weird hand thing. I'd like to see some QoL stuff like this baked into a base model so we don't have to load up a bunch of correction loras.
Yeah with base LTX I have definitely noticed similar behavior, especially squatting barefoot characters' feet turning into monkey feet with finger-toes lol. That's one of the instances where I have been testing this lora at a low weight like 0.3, just to try to keep already-visible feet from deforming into hand-like objects with I2V.
But unfortunately, conversely, this lora may have the opposite effect. In my testing there were more than a few T2V examples of hands looking like weird foot/hand combos- neither a foot nor a hand but some terrible body horror mix. tbh I don't know if that's a result of LTX just having poor anatomy understanding in general, or if this lora is inadvertently influencing 'palms' as well as 'soles', or a combination of both those factors.
The results of this lora are promising, but nowhere near a true 'fix.' I still think the only real improvements are going to come from the base model and the tooling that is only available to a real ai lab. If I can get 'decent-ish okay' results like this from 20 hours of training on 154 foot clips at rank 32 on a 5090, then presumably Lightricks could easily make real meaningful improvements to the base model's understanding of human anatomy. And I'm hoping that is their direction for future releases.
You finally did it as some of us requested. Thank you so much!
Thank you, you're a hero. This should be one of the very first LoRa for LTX. It's crazy that a big video model doesn't know anything about human feets/soles. This would be the first LoRa I'd make if I had the hardware. Now all that's missing is a foot worship LoRa, like licking, kissing, smelling at feets. That would be a dream. Thank you very much for ur work!
yep. those are feet. quite nice.
first try got 6 toes on each foot. but hey, more to love.
Create your starting image with QIE 2511 using this Lora and prompt:
丝袜足底-stocking-feet {flux2-klein-9b-base} {qwen-2511-edit} - 丝袜足底-stocking-feet-QWEN | Qwen LoRA | Civitai
I changed the original prompt to barefoot instead of stockings.
（双腿举起，赤足展示足底）模特优雅地坐着，双腿笔直向前高高抬起并略微分开，双脚完全朝向镜头，足底正对观众，脚趾自然舒展，光洁的双足清晰勾勒出脚掌轮廓与脚趾细节，脚趾圆润。整体姿态柔美而富有张力，突出腿部线条与足部美感。
T2V doesn't really work for the sole-focused from a prone/ankle extended position. Prolly just not in the pose set, but would appreciate it lol
yeah the training data only had 2 clips of 'upside down soles', if that's what you mean. I very well may add this in future, I already know I want to add squatting/tiptoe soles, which would likely be a lot of similar data as laying-prone soles.
Nice work, are you planning on making it better for GGUF quantized models in later versions? I don't know if there's anything in particular that can be done for that, since I've never trained any LTX LoRAs since I only have 12gb VRAM, I'm not well versed in video genning.
Also I'd like to see the dipping only LoRA.
THats quite perfect!
I would love a trampling lora. Girl trampling on head and body and all such things can you do that? :)
Im using this along with your footjob lora in 10EROS 1.4. I just cant seem to get past toes turning into fingers 9 times out of 10. This is definitely better than vanilla 10EROS checkpoint with DR34ML4Y lora, but LTX really needs to step up its feet game.
HIDEO KOJIMA must see this lora
Would love to see a version of this for Minimax H3, ai-toolkit already supports training on it but I would wait a few days so any bugs get fixed. Looking forward to seeing what else you put out with it!