DIS VS cloth-segmentation

Hello, I had some succes using this repo : https://github.com/xuebinqin/DIS

DIS

Rembg only adds to my frustration. It's a half-assed solution providing sub-par results and acting as a band-aid for the fundamental absence of any transparency handling in diffusion models. DIS is better, but it's a pain to set up and still often needs retouching. Trying to remove backgrounds post-generation will always be an uphill battle because the information just isn't there.

also try to look into https://github.com/xuebinqin/DIS cuz it may help to either achieve the result straight away, or to make a dataset for layering lora if each picture takes to much touch up

Also checkout https://github.com/xuebinqin/U-2-Net. They have a new version in this repo: https://github.com/xuebinqin/DIS

If you want some tricks to make this smoother, try masking the wolf upfront in each frame and then only replacing the wolf itself with the pixels from sd on the second frame and onward. Keep the first frame's sd output for the background stable or change it less quickly. I've found good results segmenting with this, but lots of good options on foreground segmentation are available.

For instance, click "open in colab" at the top of this Image Segmentation notebook

u2net_cloth_seg (download, source): A pre-trained model for Cloths Parsing from human portrait. Here clothes are parsed into 3 category: Upper body, Lower body and Full body.

Since you want to do semantic segmentation with a U-net, I would recommend tearing apart this project: https://github.com/levindabhi/cloth-segmentation