CtrlNet MiDaS Settings (JPS)

MiDaS is the depth estimator that shipped with SD 2.0's depth2img and was the default depth ControlNet preprocessor for most of 2022–2023. It's since been passed by sharper models (Depth Anything v2 is the current default most people reach for), but it's still around, still fast, and still the settings this node's own parameters are built for - the two knobs unique to this node, midas_a and midas_bg, are MiDaS's own tuning parameters, not generic ControlNet settings. This node is the control panel for a depth pass: which image to read, how hard the condition applies, when it's active, and how the MiDaS preprocessor itself behaves. It doesn't run depth estimation - that's a separate preprocessor node - it just holds the settings for it.

The inputs that matter

• midas_from - "Source Image", "Support Image", or "Support Direct". The companion Pipe node outputs this as a plain integer rather than the label text, which matches the select input on the pack's own Image Switch - it's built to drive one directly.
• midas_strength (0–10, default 1) - ControlNet weight. Depth conditions on relative spatial structure rather than exact edges, and the community-standard band for depth ControlNet sits around 0.4–0.7 for loose composition guidance, 0.8–1.1 for strict architectural adherence - again, the node's own 0–10 range is far wider than you'll normally use.
• midas_start / midas_end (0–1, defaults 0 / 1) - the step window. Ending the condition around 0.4–0.6 rather than running it the full sample is the standard advice: lock the spatial layout early, let detail generation happen unconstrained afterward.
• midas_a (default 6.28, range 0–15.71) - MiDaS's own tuning parameter used in its depth-to-normal-map math, expressed in radians (6.28 ≈ 2π is the library default). Leave it alone unless you know you need to change it; it's not something most workflows touch.
• midas_bg (default 0.1, range 0–1) - the background threshold, controlling how aggressively distant/background regions get flattened out in the depth map. Raise it if your background is bleeding detail you don't want influencing the generation; lower it if the model is treating too much of the frame as flat background.

Output: midas_settings, a BASIC_PIPE - JPS's own bundle (a plain tuple, not Impact Pack's object despite the shared type label), meant to feed straight into CtrlNet MiDaS Pipe (JPS).

Installing it

ComfyUI Manager: search JPS Custom Nodes for ComfyUI. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/JPS-GER/ComfyUI_JPS-Nodes.git

Restart ComfyUI. Remove any older install of the pack first. No model weights here - the MiDaS model itself downloads through whichever depth preprocessor node you're running (typically on first use), not through this settings node.

Where people get tripped up

If your depth conditioning looks wrong, check whether the issue is actually the depth map quality before blaming this node's settings - MiDaS is meaningfully softer and less detailed than newer estimators like Depth Anything v2, and the community's own comparisons have found MiDaS sharper-edged than the metric-focused ZoeDepth for ControlNet purposes specifically, but still behind the current generation of models. If you're getting blurry, low-detail depth conditioning and it matters for your work, that's a preprocessor choice, not something midas_a/midas_bg will fix. And, being one of the newer additions to this pack, it isn't listed in the README's node inventory, which predates it.