Keyword: armbinderoptional: from side, from behindhelpful: bare shoulders in negativefrom side is not super reliable.from behind usually have the character facing away, face totally not visible, but produces best visible armbinder shape.For checkpoints that already have concept of armbinder trained, it's really hard to remove the "strap" or boundary in the upper arm where usual elbow gloves would end.**********************************************************The lora itself isnt working too too good, but it's good enough for me. It's still a succeeded experiment and I feel like sharing the whole process.I wanted to train an armbinder, not the standard one where there's 2 straps that secures it over the shoulder. Rather, one that looks more like clothing and covers a good portion in front. Something like this.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c36c777f-7f31-4e80-9749-aa250f3b382f/width=525/c36c777f-7f31-4e80-9749-aa250f3b382f.jpeg" />So I started with the model I had when I was playing with MMD a few years ago.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/40c330e7-0db7-4fa0-8994-c09972e34a4a/width=525/40c330e7-0db7-4fa0-8994-c09972e34a4a.jpeg" />Made a MMD video with a .abc output file that can be imported to C4D/blender/Marvelous Designer. Import the model into marvelous designer and make the armbinder and fit it on the model.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/9f296e7a-48b8-4884-beec-808cc9ea92a3/width=525/9f296e7a-48b8-4884-beec-808cc9ea92a3.jpeg" />Then put it through the standard computing/recording/blender rendering process like what you would do for any normal MMD video. (<a target="_blank" rel="ugc" href="https://www.mediafire.com/file/3wc6ykzuz7sfb4h/armbWIP.mp4/file">link</a> if anyone is curious)The video itself is 5800 frames, i manually picked about 500 frames from the front, 500 from the side, 500 from behind, and trained my first lora on that.I got a first lora that is way overfitted(white hair,hair texture, red eyes, single ponytail, MMD model face, bottom plate/grey background boundary, etc.).<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/761a293c-aede-4285-b687-ce078372a804/width=525/761a293c-aede-4285-b687-ce078372a804.jpeg" />But with a lot of wildcard prompts, 0.5~0.7 strength and some img2img, i was able to generate about 120 images that I then used for 2nd round of training.The product of the 2nd round of training is what i upload here.It means I can now do whatever clothing design I want in Marvelous Designer, render images in blender, toss the blender renders to train lora1, then use output from lora1 to train my final lora.Basically, as long as I can replicate the design in Marvelous Designer, there will no longer be a lack of dataset.

Files with SHA256: 3d9c62e18c743a37eef2763beb04d6faea55ec7f5b762f2f8bf9983a85feb7d8