Confusion between large/base with depth anything v1/v2

[JonathannJacobs]

In the actual infer.py we can read :

Load model

if arch_name == 'depthanything-large':
    model = DepthAnything(**model_kwargs['vitl']).to(device)
elif arch_name == 'depthanything-base':
    model = DepthAnythingV2(**model_kwargs['vitb']).to(device)

This made me very confused because it's suggesting using the large model actually is depth anything v1 large while i was expecting the v2 large.
Is that correct ?

If it is correct i suggest renaming arch names for including v1 or v2 in it.

[shuiyued]

Hello,

The difference lies in the naming convention of the model key names.

Both are v2 models, but the network key names were modified during development.

In fact, you can achieve the conversion between our two models("DepthAnything", DepthAnythingV2) by using the following code:

# load large model
model_weights = load_file(checkpoint_path)
for key in model_weights :
  new_key = key.replace('backbone.blocks.0.', 'pretrained.blocks.').replace('backbone.', 'pretrained.')
  ckpt_new[new_key] = model_weights[key]

model_kwargs = dict(
            vitl=dict(
                encoder="vitl", 
                features=256, 
                out_channels=[256, 512, 1024, 1024], 
            ),
        )
model = DepthAnythingV2(**model_kwargs['vitl']).to(device)
model.load_state_dict(ckpt_new)

I will fix it—it’s clearly a misunderstanding.