0.9:
This is still i2v trained, but I have noticed with 0.9 that T2V works? Some motions seem like they are coming through on some t2v tests I did and they didn't seem like horror or anything. Slightly improved nudity and motion if you prompt for it. It shouldn't interrupt character loras as it's very weak on upper body and face data and is more of an amalgamation of body movements from many angles. Previews use this workflow. Modified it a bit here, use res_2s sampler from RES4LYFE on first pass, LCM on upscale pass, and my sigma changes and an optional audio latent load to track the output with custom sound voice or music.
Problems already identified: Undertrained, under ranked (should probably be rank 64 for these nsfw concepts) and some tagging changes with angle set ups and pruning of the dataset can be done for 1.0.
I also keep getting a visual bug with the upscaler in to larger vertical formats, if the image goes beyond 1536 vertically sometimes you get color noise at the top and bottom, not related to the lora. Anyone with a fix for that lemme know. Another quirk, or probably advantage, is that this lora has so much motion impact you can occasionally lower the compression on the PreProcess node and get some better adherence to start image, or in some cases turn it off completely and isolate the motion to only the lora. That's not usually plausible without loras, results just zoom in or freeze.
I tried to reform the dataset tagging to bind the base concept under:
doing a sexy move
The model seems to actually pick it's base concept up in context the way I wanted it. If you start a prompt with "a naked woman standing in front of the camera while doing a sexy move" it should start out with it. But you can do things like another action and dialogue, then later on in the prompt: "blah blah blah. Then she starts doing a sexy move." You can bring it in later on in context. At least that's the idea and I gave it clips to lead into that.
You can set these other keywords up in-context with something like:
"she is doing a sexy move with side to side hip movement and butt grinding."
Won't always work, depends on start image and the usual LTX2 finding a good seed and refining prompt until it works. You don't want to just have a short prompt using only these tags, you should still prompt the base model correctly and mesh the data from this into it with a normal long and descriptive prompt. All these tags do is just give a hint to what kind of motion and scene it's trying to do, but this is pretty strongly trained and rigid and if the image is complicit it knows what to do, but also may just start mixing different motions together regardless of prompt. That's the biggest issue to try to iron out. The entire set if mostly static camera as well, obviously for better data clarity.
Angle Description:
Seen from behind
Low angle static camera (crotch focus)
Facing away
From above
(all else is assumed to be front facing camera, probably don't prompt anything)
Motions:
grinding
bouncing
feeling
touching
rubbing
squeezing
dancing
Front:
a naked woman, a semi-naked woman, a girl
vagina and anus
sex, the penis goes inside the vagina, two people having sex
grinding
breast squeeze, grabs her breasts, rubbing, rubs, feels, playing with her breasts
breast bounce, breast jiggle
sexy dance, dancing in-place, bouncy dance
side to side hip movement
gyration, gyrations
Bouncing up and down
Back:
Seen from behind, facing away
Turns to face away from the camera
Looking back over her shoulder
Butt
Butt jiggle
in a doggystyle position/pose
butt grinding
twerking
shaking her butt
feeling her butt
spreading her butt showing her anus and vagina with her hand/hands
Sex: (don't expect much, needs targeted lora motion to partner with for a real output. This is almost just noise to make it aware of concepts. Also is a lot more riding on-top focused.)
explicit motion
she reaches down and spreads her vagina with her fingers
bouncing up and down while having sex
the penis goes inside the vagina
POV
man has sex with the naked woman
very long hard and stiff penis
Anal
Strength: 0.2-0.91, 0.75-0.8 seems the best zone. High strength has adverse effects. Always try it turned off to see if it's even helping, its not magic.
This one is actually still kind of a test of course. It is overfit, but also not? I thought the first one was, but it isn't, just the audio. The 768 bucket seemed like it was going to overfit first, but the 512 long bucket seems to still work to extend and smooth things out. Idk like most others I also find this model has a lot of quirks to training it. The wan2.2 one was trained like this and works fine, but it's all aligned with the same frame count, I never did a split frame count but I don't see any kind of interruptions happening on long generations. I have seen bad outputs and good outputs, but then when it's a very explicit image and I turn the lora off: nothing happens at all, lol. So at minimum you can just use this at half strength to give nude characters in erotic positions movement at least, and it will still have a lot of impact.
This is also mainly a motion support lora, if this was Overwatch or LoL, this lora is Mercy or Lulu. On it's own it will do a few subtle unique motions, but mainly just hone the model in on a single female subject. If you use it with more targeted NSFW loras, make sure you balance this lower, it has sex and NSFW data, but only enough to mesh in with more targeted loras. Do as I say, not as I do. This dataset is borderline untrainable with too much variance, but just enough to transfer a pretty large variety of motions and lifelike quality through scraping on overfitting. You're way better off training if you focus on one thing and one motion with a dataset of that thing in slightly different shapes and settings with very similar composition.
After testing base model more, I changed tagging for this back to the Wan2.2 lora + less vulgar terms for body parts. LTX2 with the default Text Encoder seems to understands anatomy in less vulgar terms. So butt, vagina, breasts, penis, testicles; it knows what parts of the body they are, just not how to fully depict or animate on them. Pussy, ass, cock, dick seem a lot less reliable.
As far as the loss of the sexier voice from the test: I'm trying to put together JOI data, ASMR, and dirty talking during sex clips to make that it's own Sexy Voice package.
V0.9 is expanded with 768 data at 121 frames, audio removed, since the audio training is way too impactful and can conflict with loras that actually need dedicated audio data tied to their concept.
Test 0.5: first rough version with less unique motions and motion clarity, but it has a way sexier voice lines.
Test 0.5 keywords:
Style: realistic - sexy.
Style: realistic - explicit.
Pussy, ass.
The rest of the motion same as 0.9, but use pussy and ass instead (weaker tokens).
Description
-512x 241f 24fps
-768x 121f 24fps
No audio.
FAQ
Comments (29)
I'm not getting good results, maybe its my lazy prompting. I'll try harder.
What kind of image/prompt? I actually need feedback like this. I want to see why and how it's failing. Also, this still suffers from the same thing with the base model where it's just a bad output on the seed and the motions bad, everything needs to be adjusted or just reroll and try a new one.
@tenstrip One of my uploads just got approved, its the one with angel wings. I also uploaded my workflow as json in the negative prompt section if you're interested. Ive also ran a few without the lora and it seems like your lora just seems to add unpredictable movements.
@fox23vang226 So one thing I'm still trying to fix is that I notice that LTX just doesn't like motion+talking. I assume its because it wants to draw lipsync easier. You usually have to do "motion, "dialogue", then motion, while motion: "dialogue." One thing too is the PreProcess strength if it's too low, even on 33 default, it makes the effect stronger where it seems like all the motions is pent-up and then releases at the end in a mess. Try 34,35,36 for that.
My prompt for that image might be something like "she is sitting with her legs apart. She slowly opens and closes her legs while doing a subtle sexy move. While lifting her arms above her head she says "come fuck me baby" (or something). Then she does hip gyrations and moves her hips in a sexy way." Reroll until good.
@tenstrip I guess if you use lazy prompts and dont define or detail the exact movement you just get random nonsense moves. I tried a less lazy prompt and it worked ok.
@tenstrip Im lazy though, Id prefer not to go to an LLM each time i need to prompt for a new movement. Hopefully the lora gets better with sexy movement so it kinda just forces a sexy movement so you dont have to go to an LLM every time. Kinda like how Grok video understands sexy movements without you really having to prompt really hard. Kinda what I was expecting from this Lora, to be lazy with it.
@tenstrip btw thanks for the PreProcess strength tip, im noticing a lot more effects going on in my videos after using 36. Like background people were sometimes frozen at 33, but at 36 they all move now.
@fox23vang226 Yeah very important. Sometimes you can even make it 0 if you're using a very weird or unique figure and don't want things to change.
i love the examples you posted for this lora, however my examples are not as good as yours. i was trying to get a woman to gyrate her hips on a dildo in a fast circular motion, but the dildo turned deformed and the movement looked bad... i don't have any examples, but i do have the prompt i can share if you can help me with this! i just wish i can get a good result out of this... been waiting for something like this for a while and I'm not good at prompting either...
Something for that would be: "A naked woman seen with her legs open and her vagina is visible facing camera with a dildo inside of it as she is doing a sexy move while grinding on the dildo. The dildo goes inside the vagina and she bounces up and down on it." But the actual data for that is on the weaker side-you'd want to use the NSFW lora as well and try them both at 0.5 strength and tweak it from there. Prompting is very important for this model it can really save an output. You might have to describe the dildo as "rigid and glued to the floor" and really if you see weird thing happen: change the prompt and tell it to do the opposite, that can work sometimes.
@tenstrip ahh thanks for the help, i will try that
@TypicalLoraLover It also may be futile as well, this is still a WIP lora and looking at training repeats now: I think it's fairly under trained. I'm gonna trim it down and add more dildo action instead of the clips that aren't doing anything and aim for a way stronger train for v1.
@tenstrip I am looking forward to that! Would love to see some dildo grinding action!
changed the game with this one
Trying to just get others to see the potential in the model. With some loras and merges eventually you get some lora mixes working and it becomes much better and more reliable.
@tenstrip youre doin gods work my friend🫡
Sadly I can't get it to work, it changes the model of my img2vid and it doesn't listen very well to the comments I give it.
Yeah it still seems very reliant on the base model getting in line with the prompt and a good seed. It just kind of enhances the base model a bit it's not strong enough to actually sway the base model into changing the output enough because it's still really just a 'style' of motion. I might give it the beans and go for an insane step count train and see if I can actually just cook it on there.
This seems to work pretty great doesn't seem to have the strong style influence last one did and most motion (a bit jittery) work well even on anime stuff. One thing there does seem to be a issue is breast grabs/rubs, It's not only the lora but ltx in general will just move the hand to the stomach if there is a stomach in the picture lol. Anyway thanks for lora, hope you keep at it!
Oh and also this does work well to give an image a jump start if you are having trouble with ltx animating stuff, try throwing this into it, even for non sexy stuff (lot of jiggles everywhere though lol)
I'm noticing the breasts tend to jitter a lot yes. I think it's related to the breast bouncing data which looks like I have to either guess that this one is over or under trained, but leaning towards under. These LTX2 loras look like they need 50-60+ repeats to really fluidly grasp the dataset so steps need to be set for that, where this is just about 40. I'll probably move those breast bouncing clips out of the 768 bucket because that's probably causing that, even though they were really good motion examples that I still need to find a good output to demonstrate.
One major thing about usability with the upscale workflows- if your first stage is rendering lower than 512 on a side you're going to have way reduced grab on the start image since these are close cropped data points. I'll look at going 256 bucket with it maybe that can help but try adjusting the downscale to be at least 512 width.
Hello and thanks for your works, sadly the render quality isn't the same as wan 2.2 for LTX 2. Atm i'm using a 3090 so a wan is slow xD. I don't know i used every best settings i saw and quality didn't really change. I'm pretty disapointed by LTX 2 atm. If u guys have some idea please tel me
Try res2s sampler with the 2part upscaling workflow and use the fp8 model with the distilled lora. Lower the strength of the distilled lora only for the second pass that's been the best visual quality average I've seen.
@tenstrip Hello i'm using the workflow u sent in your description i'm trying your advice atm. Thanks for your response sir :D
@dayman02400741 It's still changing a lot now. They just released new nodes and I'm seeing what they can do.
@tenstrip thanks you sir have can you good feedback about theses nodes ? Sorry to respond so late i'm testing again and again this models :D
@dayman02400741 They produce much better audio on the first pass and skipping the model's layer 29, or layers 29,28,27 produces much more motion from loras in the first pass video. The audio is also so much better since you can make the audio CFG a bit higher separate from the video. I ended up making a vae audio decode of the first pass audio, and then encode the audio directly into the second pass to force it to use it exactly instead of just making it worse by passing the latent directly. I'll post and updated WF soon with all that.
LTX 2.3 VERSION, please!!! Thanks!!