🌆DepthFM ModelLoader

Every DepthFM workflow starts here, and nine times out of ten when someone reports "DepthFM doesn't work," it's because they skipped the manual model download this loader depends on. There's no auto-fetch - the ModelLoader just lists whatever checkpoints it finds in ComfyUI/models/depthfm/ and lets you pick one. Empty dropdown? You put the file in the wrong folder.

The single input is depthfm_model, a dropdown populated from that models directory. Its one output is model (type DepthFMMODEL), which is exactly what both 🌆DepthFM and 🌆DepthFM Literative expect on their model port. Wire the loader's output to either one and you're set - there are no settings to fiddle with here, which is the whole charm of a utility node.

Worth knowing what it's actually doing, because you'll hit it when things break. It torch.loads your checkpoint, builds the DepthFM UNet from the weights baked inside, and pulls the Stable Diffusion 1.5 VAE from HuggingFace (the vae subfolder) on first load - so the loader needs internet once even though the big file is local. Then it moves everything to CUDA and locks it in eval mode. Note the hardcoded .cuda(): no CPU fallback, no Apple Silicon MPS. If you're not on an NVIDIA card, this pack dies right here with a clean CUDA error, so don't go chasing deeper symptoms.

Install is the pack's shared routine - ComfyUI Manager, or git clone https://github.com/ZHO-ZHO-ZHO/ComfyUI-DepthFM into custom_nodes, then pip install -r requirements.txt (just torchdiffeq), then restart. After that, the step that actually matters:

mkdir -p ComfyUI/models/depthfm
wget https://ommer-lab.com/files/depthfm/depthfm-v1.ckpt -O ComfyUI/models/depthfm/depthfm-v1.ckpt

If the dropdown stays empty, that's almost always a path problem. The pack registers models/depthfm as its search directory, so the file has to live there - not in models/checkpoints, which is the usual guess. The official file is a .ckpt; if you found a .safetensors conversion somewhere it'll probably load too, but the README's file is the one everyone actually runs.

It's a utility node, so that's the whole job: put the file where the dropdown can see it, and it hands your depth models to the inference nodes. Get that wrong and no amount of fiddling with steps will save you.