Include BiRefNet as a model for RemBG

[yo1nkers]

https://github.com/ZhengPeng7/BiRefNet model performs extraordinarily well for background segregation, would love to see it in Auto1111's RemBG extension by default.

[ZhengPeng7]

I'm the author of the BiRefNet. Here is also the online demo. If you think it's good and would like to include it, please let me know anytime.
I've been mainly quick to respond to all issues, and I'm glad to see more people using BiRefNet, which is 100% open-source and commercially available.

[dimitribarbot]

Thank you both for your commitments. Thank you Zheng for your hard work in segmentation model production, and thank you dimitri for working on the easy inference extension

You're welcome 😃.

It would be good to see this extension in the BiRefNet repo & extensions list. This might be a new fan favorite image segmentation for the community.

Since Automatic1111's RemBG extension was developed by AUTOMATIC1111, I thought it would be better to enable BiRefNet in the official extension instead of adding a new one. However, I don't know if or when my PR will be merged in rembg. That's why I created the unofficial extension. I'll wait a few days to see and if I don't get a response then I'll make a request to add it to the official extensions list.

[dimitribarbot]

I've just found a notable difference between rembg and my unofficial extension:

• rembg is using onnx files and as such, the models will only be inferred on images with size of 1024x1024 (@ZhengPeng7 correct me if I'm wrong). All input images will be resized to 1024x1024 before inference and then the computed mask will be resized back to the original image size.
• whereas the unofficial extension is using the pytorch versions which allow inference on any image size (width and height will be floored to the nearest multiple of 32 but that's all).

In the latest version of the unofficial extension (update to the latest version if you want to try), I've added a "Resolution" field to be able to see the difference. For instance with the following image:

When using the "General" model and original image size 480x360 (image will be resized to 480x352 to fit the multiple of 32 condition), we have the following result:

Whereas with a fixed size of 1024x1024 (what we will have with rembg), we get:

As the size is inherent to the onnx model file building, I don't think we can change the resolution when using rembg whereas it is (now) possible in the unofficial extension.

N.B. In the unofficial extension, I've also fixed an issue with the "General-Lite" model which was not working. It should be OK now.

[ZhengPeng7]

Yeah, BiRefNet can infer with a resolution of times of 32. But 1024x1024 is the suggested value, since it's the resolution used in training. Other resolutions are more likely to produce unexpected results.

[ZhengPeng7]

In this case, 1024x1024 causes some problem, but I think its still better in more general cases.

[dimitribarbot]

I've created the following PR in the Automatic1111 RemBG extension repository. @yo1nkers I've followed your suggestion and added a radio button allowing to select between Classic models or BiRefNet ones:

Let's see if this PR will be merged. Otherwise I will ask for my unofficial extension to be added to the list of Automatic1111 official extensions.