HOME | DD

Description

As people know, I've done the occasional armwrestling image. This isn't easy in Stable Diffusion - you've basically got to use controlnet to make it work, which is a lot. However, when I saw that Bing Image Creator was creating some really nice images from people like sbanon gandalfi2011 and thormanoftunder I wondered whether Stable Diffusion could pull it off with a Lora. The Bing generator is great, but you get a lot more flexibility via SD.

So I trained one using Civitai's lora training and around 100 images, and it turns out it actually works surprisingly well. The images on this post are just from the Lora, with no other loras, and a little post work and inpainting for expressions and sweat.

You can download it from here.

Limitations...

* Due to the dataset I quickly grabbed, it tends to generate mixed armwrestling matches. It can generate woman vs woman or man vs man but you need to work the prompt.

* It tends to make pretty buff people, once again, due to the dataset.

* It tends to generate almost always matches at the half way point.

While I do want to do a second one at some point, this will do for now. I'll personally likely just use inpainting to change the size of people, which is much quicker than the painful controlnet method I was doing before.

As said, you can download from here , and would appreciate reviews or buzz over there. I need buzz to do training, so it'll all go into there.

Here's the image without the header, and another one I liked.

There's also a lot more in the gallery on the Lora page.

Description

As people know, I've done the occasional armwrestling image. This isn't easy in Stable Diffusion - you've basically got to use controlnet to make it work, which is a lot. However, when I saw that Bing Image Creator was creating some really nice images from people like sbanon gandalfi2011 and thormanoftunder I wondered whether Stable Diffusion could pull it off with a Lora. The Bing generator is great, but you get a lot more flexibility via SD.

So I trained one using Civitai's lora training and around 100 images, and it turns out it actually works surprisingly well. The images on this post are just from the Lora, with no other loras, and a little post work and inpainting for expressions and sweat.

You can download it from here.

Limitations...

* Due to the dataset I quickly grabbed, it tends to generate mixed armwrestling matches. It can generate woman vs woman or man vs man but you need to work the prompt.

* It tends to make pretty buff people, once again, due to the dataset.

* It tends to generate almost always matches at the half way point.

While I do want to do a second one at some point, this will do for now. I'll personally likely just use inpainting to change the size of people, which is much quicker than the painful controlnet method I was doing before.

As said, you can download from here , and would appreciate reviews or buzz over there. I need buzz to do training, so it'll all go into there.

Here's the image without the header, and another one I liked.

There's also a lot more in the gallery on the Lora page.