MiDaS Model Loader

MiDaS Model Loader loads a MiDaS depth model as an input for the suite's MiDaS Depth Approximation node - the first step in getting a depth map for anything else in the pack that needs one, most notably Image SSAO, which needs a depth map to fake its ambient-occlusion shading.

What MiDaS actually is

MiDaS is Intel ISL's monocular depth estimation model - give it a single 2D image, no stereo pair, no depth sensor, and it estimates relative depth: which parts of the scene are closer, which are further, without claiming to know actual distances in meters. It was the default ControlNet depth preprocessor for most of 2022–2023 and shipped as the backbone of Stable Diffusion 2.0's depth2img feature. It comes in a few variants trading speed for quality - DPT-Large gives the cleanest, sharpest depth gradients; the small variant is fast enough for quick previews but noticeably softer.

It's also, genuinely, still a solid choice today rather than a relic - a well-cited 2026 community A/B test pitted MiDaS against the theoretically more accurate ZoeDepth specifically for ControlNet use and landed on "use MiDaS," because ZoeDepth's metric precision came at the cost of the sharp edges ControlNet actually needs to follow well. If your only exposure to depth models is through comparisons calling MiDaS "the old one," that's not really the right frame - it's older, but it's not worse for this particular job.

Where it fits

Load the model here, feed it into MiDaS Depth Approximation to actually run inference on an image, and that depth map is what Image SSAO wants as its second input, alongside your source image. If you already generate depth maps elsewhere in your workflow - most photoreal ControlNet setups do, usually through comfyui_controlnet_aux, which bundles MiDaS alongside Depth Anything, ZoeDepth, and Marigold in one pack - you may not strictly need WAS's own loader at all; you can often reuse that existing depth map directly as SSAO's input instead of running MiDaS a second time.

Installing it

Through ComfyUI Manager: search "WAS Node Suite", install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui
cd was-node-suite-comfyui
pip install -r requirements.txt

The model weights download automatically the first time you actually load one, so give the first run a minute and make sure the machine has internet access at that point.

Troubleshooting

If the loader hangs or fails on first use, it's usually that initial download - check your connection and console output before assuming the node itself is broken. Beyond that, the pack-wide caveat: WAS Node Suite has had no active development since December 2023, so a ComfyUI core update can occasionally break the whole suite's import over an unrelated pinned dependency. If MiDaS Model Loader is missing from your menu entirely rather than erroring at use, that's the suite-wide problem, not this node.