Female Muscle Growth
Introducing another expansion/growth LoRA. This one does not work very well with pure text-to-video, in spite of being compiled against the T2V base model. I think it's an issue of prompting, as there's no good way to describe a person as "not muscular" without triggering muscular bodies by accident, which removes much of the point of the growth model. Instead, I use this LoRA in image-to-video workflows, either starting with static starting images or by extending a video using Wan 2.2 with Vace Fun. My workflow for all my samples is embedded in the sample videos and is a workflow I called "straight to video" in my comfyui saved workflows. It generates an image using a static image prompt, then it applies that image in an image-to-video flow with additional motion/action prompt text. I found I get better results if I re-apply the static image prompt in the video prompt since it reinforces the clothing and scene.
As with my other expansion LoRA, this is just a High Noise LoRA because that's all that seems necessary, leaving you open to apply low noise models to influence the final look however you like.
For prompting, I've found a couple of things that may help you:
the keyword is "msclgrw" ("muscle grow" with the vowels removed)
explicit descriptions about how the woman looks before and after will help. Wan will do its best to render what you describe, so whether you want "athletic fitness model" or "champion bodybuilder", you have to specify that kind of thing. See the prompts in my samples for examples.
sometimes the woman's sleeves or long pants will disappear instead of simply stretching out. This happens in the Breast Expansion LoRA also, and you can sometimes prevent it with prompting guidance, and sometimes there isn't anything you can do to change it. Sorry, it's not my data set, I swear. For prompts, I usually include a phrase like "inside her clothes" to get better results. For example, "her muscles are slowly growing bigger inside her clothes", and then describe the clothes being stretched by the new body, and so on. Reinforcing for Wan the idea that the clothing should stay there, just wrapped around a new person.
I think this LoRA works best when mixed with extended video clips made with Vace Fun since five seconds isn't really long enough to tell a story about the transformation. If you decide to do that, I recommen using muscular LoRA for the video extension since it can help guide the renderer to know how her body should look when it moves around. Remember that in image-to-video or video-to-video, Wan doesn't know about anything that isn't visible in the original source, so if you can give it hints about how things look that will help. For example, if you start with an image and then have the woman turn around and face the other way, then Wan has to invent all of that, and it could sometimes use some help.
I hope you all enjoy this LoRA and please post your creations to this model page so I can see what you're coming up with. I look forward to checking out what other people can do with this model!