Krea2-UniDepth Calibrate Guide

Krea2-UniDepth Calibrate Guide is the pack's standalone depth fixer: per-image percentile clipping, gamma, explicit polarity, and optional edge-preserving smoothing. It does exactly one job and doesn't touch geometry - what goes in at one aspect ratio comes out at that same aspect ratio, just with the useful depth range pulled out of the weeds.

The honest headline, though, is that this node is legacy in its own pack. Every one of these controls is built into the Krea2-UniDepth Conditioning node behind a calibrate toggle, which is the path the author recommends for a real generation. So when do you actually reach for this one? When you want to preview or audit a depth map before committing, or when you're producing a cleaned guide for something else entirely and don't want to stand up the whole conditioning stack. It's a utility node, and it knows it.

Why raw depth maps need this at all

Depth preprocessors like Depth Anything output a map where pixel brightness encodes distance - white near, black far, per the standard convention. Two things go wrong in practice. First, outliers: one blown-out highlight or shadow can flatten the whole dynamic range, so the actual scene structure occupies a thin slice of 0-to-1. Second, the contrast curve may not suit the model, or the polarity may simply be backwards for how your LoRA was trained.

The fix is deterministic and per-image. The node converts to luminance, optionally runs a self-guided edge-preserving filter, clips to your chosen percentiles, normalizes the useful range, applies gamma, and handles polarity. The guided filter deserves a mention: it uses local mean and variance rather than an ordinary blur, so it kills noise and banding while keeping strong depth boundaries intact. It's off by default - enable it only when you actually see noise or banding.

The inputs that matter

• low_percent / high_percent (defaults 1 / 99): the percentile clipping. The README's starting ladder is 0.5/99.5 for gentle calibration, 2/98 for stronger contrast recovery.
• gamma (default 1.0): below 1.0 lifts darker depth levels; above 1.0 suppresses them.
• polarity (keep / invert): flip this first, before anything else, if your depth looks reversed.
• smoothing_radius / smoothing_strength / edge_epsilon: the denoiser. For noisy or banded maps try radius 2–4, strength 0.25–0.5, epsilon 0.005–0.02.

The single output, calibrated_depth, is a plain IMAGE, so you can route it to a preview or SaveImage and actually see what the Conditioning node will encode. That's the real use of this node: as the "guide" its display name promises.

Install and troubleshooting

It's part of the pack, so one install covers it:

cd ComfyUI/custom_nodes
git clone https://github.com/cicalooo/ComfyUI-Krea2-UniDepth

Restart, and it appears under the Krea2-UniDepth category alongside the four others. No model files are needed for this node - it's pure tensor math - but if you're using it as a stepping stone into the depth workflow you'll want the pack's functional depth LoRA in ComfyUI/models/loras/. The rest of the stack (Krea2 base, Qwen3-VL 4B encoder, Qwen Image VAE, ComfyUI 0.29.2+) is shared across the whole pack.

The classic failure is expecting this node to rescue a map that's fine but whose polarity is wrong in the model's eyes - calibration is only as good as the convention you feed it. Check polarity first, then percentiles, then gamma, one family at a time on a fixed seed. That ordering is straight from the README and it saves you from the everything-at-once spiral that makes depth work feel hopeless.