Simple LoRA to help with adjusting a subjects muscle mass. You can swing it both ways pretty far out from -3 to +5 without much distortion. Positive gives them more muscles. Negative gives them less muscles.
One issue you will find is that the more muscles they get, the more tan they get, and vise versa. Bodybuilders grease themselves up with shoe polish, and muscular guys probably get more sun... Probably as simple as that. I will try to find a way to regularize it so it is less pronounced. If you use my Skin Tone Slider you can counteract that effect.
Description
FAQ
Comments (36)
I literally need this. Like badly. I've been trying to get the right amount of muscles on my char and he's waay to ripped and then THIS appeared!
Happy to help, let me know if it works out for you, if not, tell me how I can improve it.
@Ostris Hey! So far so good! Only gen a few pics but it's going quite well. I'll keep it up over the next few hours but I don't think there is any problem. This is great! Thank you!
Thank you for sharing this series of function-adjusting LoRAs. The functionalities you've achieved and the effects are truly impressive!
Previously, I too created a functionality LoRA that adjusts the amount of clothing on characters. One major issue I encountered was that when weights greater than 1, there was a noticeable overfitting problem. I sincerely wish to ask you how you manage to prevent overfitting within such a large adjustment range, from -5 to +5. I have documented my production method on the model description page of the Clothing +/- Adjuster. If it's helpful, you're welcome to reference and compare it.
Additionally, I noticed you mentioned a production concept in Nighttime LoRA. I'm not sure whether the current fat-thin and skin color adjustment LoRAs are based on the improvements to this concept. I didn't fully understand the process of this concept. If you could provide more detailed guidance and explanation, I'd be extremely grateful!!
Hi there. Thanks to your comment I found the Lora clothing adjuster and it's something else that I really needed. Thank you for making it. It so hard to get a char completely nude for Lora purposes and bam! I find this Lora and then yours in the span of twenty mins!
Thank you, I am happy to contribute to the community that has helped me so much. The sliders I kicked out today use a different method than my seed locking method described in the Nighttime LoRA. These are not trained on images, instead it is pure latent loopback training. It is loosely based on https://github.com/rohitgandikota/erasing . I have been using some of the work from https://github.com/p1atdev/LECO as a reference, that is a great place to get started on the concept. But I am not erasing things as much as transferring their activations to other concepts. The general idea is, you take a set of concepts like "scrawny", "average build", and "muscular, run them through the text encoder and process them similar yo hou you would normally train with a noise scheduler and try to predict the noise, only we use the "average build" term as normalizer, and try to reduce the prediction of the "scrawny" to it, while also adding the difference to "muscular" loss to amplify it, giving us a divergence between the two concepts.
It still needs quite a bit of work with normalizing the training to prevent forgetting other concepts. I have none of that currently. I train it for about 1000 steps, and only do 8 dim LoRA on most. I have gone as low as 4 with great success. Since the model is only training on itself, and the rank is so low, there is not much to over fit to, though I am sure it can happen.
When training on actual data, I frequently overfit, even when doing seed locking, which shouldn't happen in theory, and maybe it is just me doing something wrong, who knows.
@Ostris I expected to it to be something simpler like training for one concept normally, then inverting the weights and training for the opposite concept, and back and forth. I must look into LECO.
@Ostris Thank you for clarifying my doubts. I noticed LECO a few days ago, but I didn't expect that the variants inspired by it could be so stunning! It indeed simplifies the process compared to the technology I'm currently using, and it is a great way to handle concepts already present in the base model. I look forward to you publishing more works based on this method! I also hope that one day, when the method is more mature, you can open-source it on GitHub. The emergence of more functional lora will invigorate the entire community 😄
me on the left
This is extremely useful. Do you intend to train a LoRA based on body hair slider in the future?
That is an amazing idea! I will see if I can get one done in the next few days.
@Ostris Yay thank you :-)
@Ostris Hair styles could be useful as well though it's not a slider per se. I find it so hard to get innovative haircuts on male subjects.
You just read my mind. When I saw that weight slider I was thinking about it
Super excited to play with this, I've been battling everything in all of your new LoRAs with prompts but these details all seem baked in so this should be super useful!
Love this one!
This is incredibly useful, especially since many models tends to make males incredibly ripped all the time.
The only downside is that low weight tend to also make the subject younger so it's difficult to have actually skinny adults, but still great!
The deaging should be fixed on the next iteration. Hopefully.
Love it very much! Thanks for sharing. No need to describe all again.
So the weights basically move in a certain type of direction in some multidimensional space, right? So I guess there's always a trajectory between various weights and various concepts.
So a Lora like this will move the weights, let's say like 200 different weights, in a certain direction, right? But realistically it couldn't be linear, could it?
Like, there's not a linear transformation between weak and muscular as far as weights go?
It is trained while continuously flipping the network back and forth to create a binary understanding between the two concepts, making the weight transition linear. As it is not amplifying everything that would make someone muscular, in this case, but instead amplifying the difference between the concepts of muscular and the negative, which was the concept of scrawny. In reality I am using much more than one concept per side to balance it, but that is the general idea
@Ostris so I actually kinda like how the guy on the left, you can see his ribs. But I also want the pectorals defined and abs. I wonder if turning muscles up on your slider would make it impossible to show ribs like that.
@Janet It may work if using in conjunction with my weight sider. You should be able to make a character muscular and emaciated at the same time giving them a very lean muscular look like that. I haven't tested, but in theory it should work
HELP!!!!! RuntimeError: The size of tensor a (2048) must match the size of tensor b (768) at non-singleton dimension 1
It sounds like you are running it on an incompatible model. This will only work on sd1.5 based models
i have a question,if i make two characters in one picture, will this lora affect all of both?
Yes, it will affect all people in the picture. You can isolate them with inpainting if needed, that is what I do.
This is the best muscle model around that is compatible with a realism base. At least from what I've found. I hope you do another one eventually with even more data.
Sorry, dumb question, but how do you actually install sliders into Stable Diffusion?
When the type here is listed as lora you install it in the lora folder. Do a google search for "stable diffusion lora".
@LumpyOatmeal Oh so it's like the other Loras? And when it says "sliders" it's just something like (muscle:1.2)?
@alferitu2858 that would be my guess, using a weight in parenthesis.
put <lora:muscle_slider_v1:-1> in your prompt.
Best results for changing muscle mass for me was <lora:muscle_slider_v1:-1> (less) to <lora:muscle_slider_v1:2>(more)
hello, thanks for this LoRa, but I need more descriptions to use this. Basically I upload an image of mine and try to look more muscular. Would you please provide some sort of more detailed guide? thanks
This is a lora. You need to put the lora into the prompt
If you want to change a person you can try with inpaint.
But this will not look good.
I think its working better with txt2img
<lora:muscle_slider_v1:-3>
<lora:muscle_slider_v1:-1.5>
<lora:muscle_slider_v1:5>
are these Slider Loras a lot harder to do for SDXL? They work so well
Details
Files
muscle_slider_v1.safetensors
Mirrors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
肌肉滑块_v1.safetensors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
1a58e047-515a-4e7e-8865-30fd49473937.safetensors
maskels.safetensors
muscle_slider_v1.safetensors
muscle_slider_v1.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.