Question about `sd_locked` in ControlNet 1.0 and 1.1

[xyfJASON]

In the paper, it is said that the parameters of Stable Diffusion(SD) are locked when training ControlNet. However, when I looked into the released ControlNet 1.0 checkpoints, I found that the weights of SD UNet decoder are different from SD1.5 (v1-5-pruned.ckpt). I notice that there's an option sd_locked that enables training SD UNet decoder. So does this mean that ControlNet 1.0 are trained with sd_locked=False?

Furthermore, different from ControlNet 1.0 checkpoints, the ControlNet 1.1 checkpoints don't contain SD weights and need to be used along with v1-5-pruned.ckpt, does this mean that ControlNet 1.1 are trained with sd_locked=True?

Here is the script I used to compare two checkpoints:

import torch

ckpt = torch.load('./ckpts/v1-5-pruned.ckpt', map_location='cpu')['state_dict']
ckpt2 = torch.load('./ckpts/control_sd15_canny.pth', map_location='cpu')

notclose = []
maxdiff = 0.
for k, v in ckpt.items():
    if k.startswith('model') and k in ckpt2:
        if not torch.allclose(ckpt2[k], ckpt[k]):
            notclose.append(k)
            maxdiff = max(maxdiff, torch.max(torch.abs(ckpt2[k] - ckpt[k])).item())

print('notclose:')
print('\n'.join(notclose))
print(f'maxdiff: {maxdiff}')