☁️BizyAir MiDaS Depth Map

MiDaS was the baseline depth model for most of the SD 1.5 era, it's what shipped inside Stable Diffusion 2.0's depth2img, and it was the default ControlNet depth preprocessor before anything better came along. Depth Anything has since taken over as the community's general-purpose default, but MiDaS is still perfectly serviceable, and worth knowing about specifically because of a counterintuitive finding: MiDaS actually beat its metric-accurate successor ZoeDepth in a widely-cited community A/B test, because sharper edges mattered more for ControlNet conditioning than real-world scale accuracy did. Fast, reliable, and not the outdated choice its age might suggest.

The controls

a (default ≈6.28, i.e. 2π) is the angle parameter used when this node derives a normal map internally as part of its processing, in practice you rarely need to touch it for a straight depth-map pass. bg_threshold (default 0.1) sets how aggressively background regions get separated from foreground in the depth signal, raise it if background clutter is bleeding into the depth map in a way that's confusing the downstream ControlNet. resolution is the usual preprocessing-size tradeoff, 512 by default, push it toward 1024 if your final generation is large and needs the depth map to hold detail at that scale.

Output is a single IMAGE, a standard grayscale depth map (closer is brighter), ready to feed straight into a depth ControlNet loader.

When to reach for this over Depth Anything

MiDaS's honest selling point in 2026 is simplicity and speed, not best-in-class quality. If you're chasing maximum fidelity on a complex or unusual scene, a newer relative-depth model will generally edge it out. But for straightforward compositions, portraits, product shots, typical interior/exterior scenes, MiDaS still produces a clean, usable map, and running it through BizyAir's cloud means you skip installing anything locally at all, no model weights to place in a models/ folder, no VRAM budget to account for.

Typical depth ControlNet weight guidance applies regardless of which preprocessor produced the map: 0.4-0.7 for loose guidance that leaves room for the model to add its own detail, 0.8-1.1 when you need the spatial layout held tightly, and ending the ControlNet's influence partway through sampling (around 40-60% of the way) rather than running it the whole way tends to give cleaner results either way.

Installing it

Through ComfyUI Manager: search "BizyAir" and install, then restart. Or via git:

cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git

You'll need a BizyAir/SiliconFlow API key set up before this executes, via the "click to login" prompt the README points to on first use. Signup runs through phone-number verification, and some users outside China have reported getting stuck there.

Common issues

If depth values look muddled between foreground and background, bg_threshold is the first thing to adjust before assuming the model itself is failing. As with every node in this pack, a run that stalls or fails with no clear workflow-level cause almost always traces back to a missing or invalid API key.

Separately, worth knowing regardless of which node prompted you to look into this pack: ComfyUI Manager occasionally auto-installs BizyAir on its own when it misresolves an unrelated missing-node error, particularly on older shared workflows built before Manager's node registry improved in early 2025. It's real software from a real company, SiliconFlow, not malware, and it's safe to remove through the Manager if you didn't ask for it.