BD Normalize Luma

Every luminance-based tool in ComfyUI has the same quiet failure: it computes a threshold from your image's actual min and max, and if one blown-out highlight or one dead-black corner defines those extremes, the entire range collapses and the mask comes out garbage. BD Normalize Luma is the pre-process that fixes exactly that - it rescales an image's luminance to a target range using percentile clipping, so a few outlier pixels don't get to set the goalposts.

The core question it answers is simple: "what's the darkest 1% and brightest 1% of this image, ignoring the spikes?" From those robust bounds it remaps the image so the real content spans your chosen target_min (default 0) to target_max (default 0.95). clip_percent_low and clip_percent_high (both default 1) control how much of the extreme tail gets ignored; 0 means "use absolute min/max," which is the thing that was broken to begin with.

The two scaling modes

This is where the node stops being trivial and starts being genuinely useful:

• Range fit (default, proportional_scale off) - remaps source [min, max] to target [min, max]. Stretches both ends, lifts or drops the black floor. Good when you want full dynamic range out.
• Proportional (proportional_scale on) - a single multiplier so the brightest non-outlier pixel lands exactly on target_max, with a hard clamp above it. Input zero stays zero, shadows don't lift. This is your "dim the whole thing without touching blacks" mode - the tooltip's example is shifting the brightest part down to mid-grey with target_max=0.5 and guaranteeing nothing exceeds it, even saturated pixels.

You also get a mask input that does double duty: it restricts the percentile calculation to a region (e.g. skin only), and by default (apply_to_mask_only on) restricts the application too, so pixels outside the mask pass through untouched. That's the "balance the skin tones without touching the background" trick.

The tone-band controls

The optional luma_apply_min / luma_apply_max / luma_apply_feather trio lets you target a tone band instead of the whole image - highlights only (0.6 → 1.0), shadows only (0.0 → 0.5), midtones only (0.3 → 0.7). The feather softens both edges so you don't get a hard transition line. The blend is computed from the original luma, which is the detail that prevents feedback loops. preserve_color (on by default) rescales RGB proportionally; turn it off and you get a grayscale using the normalized luma.

Outputs: image, plus found_min and found_max - the source luma values that mapped to your targets. Those two floats are great for diagnostics; the README's BD Luma Stats node produces the companion numbers (like the recommended outer band) for wiring into shader uniforms.

Install

In ComfyUI-BrainDead: Manager search "BrainDead", or clone + pip install -r requirements.txt, restart. V3-API pack, so update ComfyUI first if needed.

The classic mistake is turning this into a fishing expedition - if you normalize the whole image when you only needed the mask region normalized, or crank target_max to 1.0 and clip highlights you wanted to keep, you've traded one artifact for another. Start with defaults (1% clip, target_max 0.95), wire the mask if you have one, and only then reach for tone-band targeting. The typical payoff: a too-dark character render whose BD Luminance Mask came out as a black blob suddenly produces a usable mask after a single normalize pass.