What is the difference with BiRefNet
He has the same model size and similar results as BiRefNet, it's hard to believe they're not the same thing.
It's the same thing from my understanding. They used BiRefNet as the base model and then they trained it with their own dataset. Hence why they achieve similar results.
The architecture used for the model is the same (kudos to BiRefNet guys).
We trained the model from scratch (random weights) on our own high quality curated training set
You can see our model exceling also in complex imagrs in which the foregroud in not a single unique object.
We have some sample cases in the model card
We trained the model from scratch (random weights) on our own high quality curated training set
In my testing, it's definitely better in some areas than BiRefNet, but one thing it sort of struggles with sometimes is mainly 2D anime characters. Specifically Where they either take up most of the image or where they are for example blending with a almost solid color background or parts of the hair that are either similar in color to the background or areas where for example a character has their hand on their hip, but while the model does remove the background, it struggles to identify the area inside the arm where it is holding to the hip as part of the background even tho they are the same color as the background. Sometimes shadows are interpreted to be part of the character too sometimes.