Scroll down and read the prompting guide for optimal results.<h3 id="goals-of-this-lora-(doesn't-mean-i-acheived-them):">Goals of this Lora (Doesn't mean I acheived them):</h3><ul><li>This Lora is intended to give more control over a character's muscle. It was trained on videos of men that I thought were very defined, regardless of their bulk.</li><li>By design the people you generate with this model should look like they have very low body-fat, I haven't tested trying to make bulky guys with higher fat like a bear body type or a strongman, my guess is it won't as it has a heavy bias towards being lean.</li><li>Other muscle focus Loras tend to bias towards very bulky, sometimes unrealistic amounts of muscle. I wanted the ability to make a guy either a very lean twink/twunk with ripped abs, an aesthetic greek statuesque natural bodybuilder or a strong roided up bodybuilder, but not an unrealistically massive one.</li><li>To be specific enough about muscle groups when captioning to be able to prompt "visible serratus muscles" or "visible veins" to have more control over muscle definition. In theory this would also work on females generated, though it was trained only on males.</li><li>I wanted to improve solo male nsfw a bit as well. Many Loras for LTX are heavily biased towards women and as a result, thinks that pecs are boobs so a lot of guys I generated with those LORAs looked like they had breasts. This isn't supposed to be a general NSFW male LORA, however I found that it does improve the look of cocks compared to sulphur base.</li><li>This is a T2V model. It wasn't trained to do inference off of a starting frame so I imagine it probably won't do I2V well. My thought process was if you have a starting image, then they already have the aesthetic you want most likely.</li></ul><h3 id="training-details:">Training Details:</h3><ul><li>Trained on Sulphur. Text encoder was Gemma-3 12b abliterated.</li><li>52 good quality videos (up to 10 seconds each) and captions, easily enough for a good Lora, but for more fine control over muscle groups I would like to up this number. There are less videos involving legs and back, I would like to add these for more control over the look of those, but this still does well with them. There are a few lower quality videos, but I tagged them as "low quality" so you can put that in your negative prompt if the quality isn't as good, I included them always because the physiques in them were great.</li><li>Big thanks to antilopax for making awesome run pods, good at responding to questions, and they actually work unlike half of them. I used his run pod template to train this.</li><li>Rank 64, might overpower the prompt, I may lower this for future runs if people and myself find it is impacting prompt adherence</li><li>2000 steps at gradient accumulation 4 so equivalent to 8k steps at grad_accum = 1</li><li>It was trained on sound, though my captions were so-so for it. I find my sound is not impacted too much by this. Some LORAs aren't trained on sound so they make videos have distorted sound</li><li>Took Roughly 18 hours on an RTX6000 Pro. Captioning by hand took 3 hours (I tried to use an LLM to do it or at least lay the foundation, most are terrible at captioning accurate at the level of detail I want.)</li></ul><h3 id="take-these-pros-and-cons-with-a-grain-of-salt-because-ltx-is-so-picky-about-prompting-that-it-could-just-be-that-the-ones-i-tried-produce-very-good-or-bad-results">Take these pros and cons with a grain of salt because LTX is so picky about prompting that it could just be that the ones I tried produce very good or bad results</h3>Good:<ul><li>Abs, it does abs pretty well, the abs are defined well and fit on the torso, it also gets some oblique and serratus definition well.</li><li>Pecs, it does a variety of them well, from slim to bulky</li><li>Abs rubbing. There are a good chunk of videos in the training where the guy rubs his abs and it makes a sound as his fingers catch on each one, this can replicate it decently well</li><li>Actual definition, not just making a big bulky guy with abs and pecs without the finer details in the muscles</li></ul>Bad:<ul><li>Take these pros and cons with a grain of salt because LTX is so picky about prompting that it could just be that the ones I tried produce very good or bad results</li><li>Audio could still be better, people sound like they're from a 70s movie sometimes, maybe it's just bad prompting on my end.</li><li>Not much training on lower leg and back muscles. These are less-seen muscle groups in videos, but I want to get more in future versions, if anyone has any suggested videos to include, even if not for back and legs, I'll take a look.</li><li>Could use more NSFW training, it would be nice to have a more general model focusing on male nsfw specifically, as there are already a lot of female focused ones. It can do masturbation and cock shaking decently, but it's not always consistent. I would need more training data on those concepts.</li><li>Idk what CFG and step size to use, I'm still experimenting, but CFG 7 and 100 steps seems to be what works on my workflow.</li><li>Not enough fine control, If you specify abs, they're likely to also have serratus and oblique muscles, as it's very common for a defined person to have all of those things. It needs more training data to fully isolate those concepts.</li></ul><h3 id="prompting-guide:">Prompting Guide:</h3><ul><li>There is no trigger word, but certain words will give you better results</li><li>"Defined eight-pack abs" will get you really defined abs, you can also try just "eight-pack abs" by itself for something lighter</li><li>"Extremely muscular man" especially combined with "bulky build" will get you a very thickly muscled bodybuilder at the end of his cut with lean muscle.</li><li>"He runs his hand up and down his abs repeatedly, they make a shlik noise every time his finger goes over an ab muscle" the lora is trained on a few ab rubbing videos and if you've seen them you know what I mean, its like the sound of running a hand over a washboard. I think the training captured some of this sound, but it's not perfect, you might want to try a few runs and see what you get.</li><li>"Skinny" will get you someone with more of a twink build. You can combine this with "defined eight-pack abs" to get a skinny twink with defined muscles. For example: "The subject is a skinny man in his twenties. He has defined eight-pack abs. He has skinny arms and slim thighs."</li><li>I captioned the training data like: "He has thick pecs. He has defined arms with veins slightly visible beneath the skin. He has thick obliques." So that is a good way to describe the person you want.</li><li>If you want body hair, put it in the prompt, if left out it'll likely default to smooth, waxed bodies. So I captioned body hair like this: "He has armpit hair. He has hair around his nipples. He has a trail of hair leading from his navel to the hem of his boxer-briefs." If you want a bear-type then try "He has hair all over his body."</li><li>I also made sure to make note of skin tone, so you can say "dark skin ", "fair skin", "tan skin", and "pale skin"</li><li>I added what I thought the approximate age of each person in the training set, so it ranges from "late teens" to "late forties". So you could say "the man appears to be in his late twenties" if you want to specify that. I did this mostly because older people tend to have a thinner and more textured look to their skin so I didn't want something like perfect skin on someone old or wrinkled skin on someone young.</li><li>To describe pecs, I used "thick pecs", "plump pecs", "extremely thick pecs", and "thin pecs". Thick and plump basically mean the same thing here, but thin is what you use for someone with an almost flat chest and a tiny bit of pec definition. "Extremely thick" is what you use when you want a huge bodybuilder type.</li><li>For more definition you can talk about seeing muscle fibers like: "muscle fibers are visible between the pecs and on the deltoids"</li><li>For veins you should use "veins are lightly visible on his X" where X is the body part, typically it was lower abs, forearms, or arms.</li><li>Self POV! This is supposed to be like those videos on social media where guys post a video of their torso but from their own perspective, so their phone camera would basically be sitting on their upper chest, pointing down at their bodies. You would say "the camera is pointed down at his own body from his point of view" or "from his perspective". I wanted this one because it's really hot, but it's not consistent right now, it kind of does it, but has a habit of giving them 20 ab muscles or multiple bellybuttons which is an annoying issue that I deal with even on the base model. I think it's an angle that LTX struggles with because it probably doesn't have much training data on it as it's an angle I've mostly seen just used in male-focused solo porn. In future versions I'll train the model more on this perspective and hopefully it will be better.</li><li>Specify shot size: "medium shot", "full shot", and "close up" or "close-up" if you prefer. A medium shot is from the hips to the head. A close-up is anything closer than that. A full-shot is a full-body shot, but those weren't common in training so you might get mixed results unless you add more qualifiers like, "the camera is high up" or "the camera is on the floor, pointed upward" or "the camera is very far from the subject". Usually if I tack those on I can get a full shot to happen.</li><li>Specify camera movement. I used the words "pan" to do it so something like" "pan right" or "pan from an overhead view to his side". You can also do "shaky handheld camera" for a phone video of themselves.</li></ul><h3 id="recommended-settings:">Recommended Settings:</h3>This is what works on my workflow, I feel like they're weird, but maybe it's a quirk of the workflow. Im pretty sure I'm using the one in the Sulphur repo.<ul><li>CFG: 7</li><li>Model Strength: 1</li><li>Steps: 70 to 100</li><li>I haven't experimented with sampler settings much but I use Euler_ancestral</li><li>Make use of the negative prompt, if I get too much bulk, I put "bulky" in the negative and it slims them down just a tad. If they're too tall or short then put the one you don't want in the negative prompt.</li></ul>Send me more training data! I'm looking for:<ul><li>Any video you think I would like</li><li>High resolution muscular guys either nude or in underwear/shorts flexing, full body showing.</li><li>Videos of defined backs and lower legs</li><li>Videos of people that have one kind of muscle showing, but not the other ones. So maybe someone that had serratus muscles without having abs (sounds rare). Or someone with huge pecs or huge arms while the rest of them is very average.</li></ul>Give me tips if you're an experienced Lora trainer. Give me feedback if you tried out the model! I appreciate it!Future Plans: <ul><li>More training data</li><li>JoyAI-echo lora. Echo is a very good LTX fine-tune model with great anatomy built in. Problem is that it doesn't seem to be trained on nsfw. This Lora and other ones didn't seem to be enough to get good male genitalia, but Sulphur can do this well enough with the right prompting and CFG. I would love to see a nsfw echo checkpoint from someone. It would probably take too much money for me to train it as I train on run pod currently and would probably need multiple H100s for multiple days. </li></ul>

Files with SHA256: dc4166001b7d02c320d962c4ae1a0d4810a8558cb74482665b5b0fd74d60a612