Adds muscle to the female protagonist of your video. The model works well with both Wan 2.1 and Wan 2.2. This LoRA can be treated like a slider with a strength range from 0.25 up to 4.0 and beyond, although there are consequences for high values that I'll describe below.
VERSION 2
Most of this text has been re-written to account for Version 2. There were too many changes in how the model behaves to just put the changes in the version notes. If you've been using Version 1, please re-read these notes as they may help.
Important: This model is 100% fictitious and is not based on images of any real person. The data set for the model was generated from still image checkpoints and has evolved from there using only AI rendering. No images (neither actual nor AI-generated) of any real person were ever involved in the making of this model.
The key phrase is "muscular woman" as in "a muscular woman is walking in the park".
0.25 - 0.75 is an athletic woman who exercises regularly and looks a little muscular
0.75 - 1.50 is an obviously muscular woman who works out and lifts weights regularly
above 1.5 is a bodybuilder level of muscular, but with you will see a little oversaturation in the body above 3.0.
For enhanced results without the penalty of saturation and less motion, you can use prompt hints. Wan 2.2 is especially good at respecting prompt descriptions for this LoRA. A simple change is to say "an extremely muscular woman", but you can go into more specific detail if you want. Beware of using the word "bodybuilder" with Wan, however, as that will likely influence the poses and motions that you get. Wan seems to know about bodybuilder competitions and posing, and those influences will appear in your video if you use that word in your prompt. Also, you should be aware that Wan understands the context of the scene. If your scene is set in a gym lifting weights, the results will be quite a bit more muscular than the same prompt and seed set in a restaurant.
This model will also increase breast size slightly, but not significantly. If you want to change the bust size more, a combination of prompting and additional LoRAs will help. This muscular LoRA does play well with others.
One last prompting hint. With Wan 2.2 and its split model pipeline, you have the freedom to use different values for high and low noise, giving you different results. Changes in the high noise pipeline will affect the overall motion in the video as well as the overall shape of the people and objects in the scene. Generally speaking, you should avoid very high LoRA values in the high noise pipe since that will hurt your video motion. Changing the low noise pipeline will influence the final look of the video with all the details. You can boost values here to get a stronger LoRA effect without changing motion, but beware this is where oversaturation can occur.
In my experience, the model looks best up to about 2.00 with very few changes to motion and saturation. You can go up to 4.00 (both high and low noise) and you will still see natural motion and clear details. Any higher in the low noise pipe and the image starts to fall apart. If you want even more extreme effects, I was able to go even higher by raising the high noise value to 5.00 and higher, while keeping the low noise at about 1.5 or 2.0 where the results are clear and not so saturated.
My sample videos include label text that shows the strength that was used for each so you can see what the effects are. Feel free to play around with the values on your own until you get the look you want.
Notes on Version 2 LoRA training: The more I used version 1, the more annoyed I was getting at the oversaturation of the faces. With this version, I used the same data set of about 200 images, but I used some comfyui nodes to automatically mask out every face with a transparent rectangle. The hair is still visible, and nothing else in the images was changed. My captions were all very simple and only had a couple of variations to counter a bias I noticed from version 1:
a portrait of a muscular woman. her face is censored. (she has her hands on her hips.) (her hands are closed in fists.)The last two sentences were optional, I added them to whichever images needed them. It seemed like too many of my data set pictures had bodybuilder poses with closed fists and hands on hips, making a lot of the output videos carry that bias. This Version 2 does not seem to have that issue. Amazingly, I am publishing Epoch number THREE out of 20. With limited captioning and a broad data set, I was amazed at how quickly and strongly Wan learned the model. With such a low epoch number, it looks like this model preserves a great deal of the motion and other details of the scene that you would see if you weren't using the LoRA at all, which is a great sign. After publishing all the samples, I'll be posting a series of videos that were all the same clip and prompt ranging from no LoRA at all up to 4.0 and a little beyond so you can see the changes in motion and detail.
I hope you all enjoy this dose of strong women.