For AMD (Directml) users

[MythicalChu]

Clone the repository and remove torch from requirements.txt (assuming you already have torch-directml installed) before pip installing it.
Delete or comment-out the line with torch.cuda.ipc_collect() from nodes.py on custom_nodes\ComfyUI-Marigold folder.

On the workflow, on the node:
reduction_method: mean, works on fp16
reduction_method: median, errors on fp16

scheduler: DEISMultistepScheduler, is the only one that doesn't freeze during the denoising step of the node.

That's how it works for me with my Rx580 (8Gb), on Windows. I only do 1 n_repeat_batch_size at 768 resolution or 2 at half that resolution, but maybe I can go higher. Half the recommended resolution is fast to compute and still gives better results than Midas, imo.

[KillyTheNetTerminal]

please can you help me? Rx580 here to, so I triying to install this custom node but Comfy just "Import Failed" I try manually and it's the same. I'm looking for torch.cudaipc_collect() too but I can't find it inside nodes.py