Normalize Depth

If you've ever plugged a fresh depth map into a ControlNet and gotten mush, the problem usually isn't the model - it's the range. MiDaS gives you one scale, Depth Anything another, Marigold something in between, and a raw estimator can spit out values from -3 to 47. Depth Normalize exists because almost everything downstream in this pack assumes depth lives in [0, 1], and most of the time yours doesn't. It's the gateway node of the whole toolkit: wire it right after your depth estimator and every other node here behaves.

How it works

The node reads your depth as a single grayscale channel (Rec. 709 luma weighting) and rescales it per image, so a batch of frames each get their own sane range instead of sharing one global stretch. Pick a method:

• Percentile (default) - clips the histogram at your low and high percentiles, then stretches what's between to 0–1. Robust to one hot pixel or one black hole in the image.
• Min-max - stretches the full min/max. Fast and obvious, but a single outlier owns the whole range.
• Fixed range - treats low and high as literal depth values and maps that band to 0–1. For when you know your estimator's absolute scale (ZoeDepth-style metric depth) and want to pin it.

invert flips white/black - some estimators output far-as-white, and downstream nodes in this pack assume near-as-white unless you tell them otherwise. gamma applies a curve via pow(x, 1/gamma); a gamma under 1 brightens the midtones, which is your fine-tune for "my foreground washes out."

Inputs and outputs that matter

Realistically you set three things: method, low/high, and invert. The defaults (Percentile, 1/99, no invert) are a genuinely good starting point - don't touch them until you have a reason.

Outputs:

• normalized_depth (IMAGE) - wire this into every other depth node: Depth Range Masks, Depth Cleanup, Depth Colormap, Depth to Surface Normal, and the 3D exporters.
• range_report_json (STRING) - per-image low/high actually used, plus method/invert/gamma. Throw it into a ShowText node if you want to audit what happened.

Install

From ComfyUI-Depth-Visualization by gokayfem (the author behind Decartunizer and ComfyUI-Texture-Simple). Easiest route is ComfyUI Manager: search for ComfyUI-Depth-Visualization and hit install. Or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-Depth-Visualization.git
python -m pip install -r ComfyUI-Depth-Visualization/requirements.txt

Restart ComfyUI and the node lives under the depth/toolkit menu. The bright side of this pack: requirements are just numpy and Pillow, there are no model downloads, and it runs on CPU tensors. No new weights, no network calls.

Where people get burned

• high must be greater than low. The node raises a ValueError if you swap them. It's the single most common head-scratcher.
• Percentile is per image, not per batch. If you normalize a video as one batch, each frame gets its own stretch, which means brightness can breathe across frames. For per-clip consistency you want a Fixed range after eyeballing the min/max from one frame.
• Don't skip it before mesh export. DepthToMesh and the parallax node both read raw values and assume [0,1]; feeding them an un-normalized map gives you geometry that's all wrong scale or nothing at all. Normalize first, export second.