Note that this model works well at higher resolution too (in contrast with MiDaS family) given the depth head is convolution and the DINO v2 base can interpolat

Note that this model works well at higher resolution too (in contrast with MiDaS family) given the depth head is convolution and the DINO v2 base can interpolate positional encoding to generate higher-resolution features: https://x.com/drawthingsapp/status/1757486115802726744 https://x.com/drawthingsapp/status/1757486115802726744