DAP Loader

DAP Loader is the gateway to the whole comfyui-dap pack, and it's the rare loader that's actually interesting. It pulls in Insta360's DAP (Depth Any Panorama) model - a monocular depth estimator trained specifically for equirectangular 360° images rather than flat photos. If you've used Depth Anything or MiDaS, the concept is familiar, but the domain is not: DAP was built to produce depth that respects sphere topology, so the seam where a panorama wraps around stays consistent instead of tearing. This node is how you get that model onto the graph.

The boring-but-critical part is that it does the download for you. On first run it looks for model.pth in your checkpoints folder, and if it's not there it pulls the official weights from Insta360-Research/DAP-weights on HuggingFace into ComfyUI/models/dap/ with a progress bar. That's the whole "Auto-Patching" the README mentions too: the upstream research code has hardcoded relative paths that assume you're running inside the DAP repo, and this loader changes into the bundled dap_core directory before instantiating the model so it doesn't blow up.

Inputs you'll actually touch:

• model_size - one option, vitl. Insta360 only shipped the official Large weights, so this dropdown is a formality. Don't hunt for a vitb or vits; they don't exist yet.
• precision - fp16 (default), fp32, or bf16. This is the real knob. The included workflow template recommends fp32, and there's a reason: the author's own testing found fp16 less stable for this model. If you're chasing speed, try fp16 first - but if the depth maps come out full of salt-and-pepper noise or NaNs, flip to fp32.

The output is dap_model, a custom DAP_MODEL type - not an image, so don't try to preview it. Wire it straight into DAP Inference's dap_model input, and that's the entire job of this node done.

Installing

ComfyUI Manager users: search comfyui-dap and install. Everyone else:

cd ComfyUI/custom_nodes
git clone --recursive https://github.com/maxious/comfyui-dap

That --recursive is not decoration. The pack vendors the official DAP repo as a git submodule (dap_core), and if you clone without the flag the loader crashes on import with a missing-module error. Then install the heavy dependencies into your ComfyUI venv:

"path/to/ComfyUI/venv/bin/python" -m pip install -r requirements.txt

requirements.txt is a serious list - open3d, utils3d, mmengine, einops, plyfile, opencv-python, xformers, plus friends. On Windows the README also recommends reinstalling torch/torchvision/xformers from the cu130 index. None of this is a one-second setup; budget a couple of minutes and a decent internet connection, because the first time you run DAP Inference it also downloads the weights.

Common issues

• Cloned without --recursive → import error about DAP modules. Fix: delete the folder, re-clone with the flag.
• First run "hangs" - it's not hanging, it's downloading a ViT-Large checkpoint and then doing a one-time model load. Watch the console.
• Funky depth output on fp16 - known quirk of this model; switch the loader to fp32 before you go down a debugging rabbit hole.

It's a single-purpose node with a single-purpose model, so there's not much else to tune. Load it, pick your precision, move on.