Policy Guard

Every delivery has a spec sheet - peak nits, how much clipping is tolerable, how much black crush, a required metadata set. Policy Guard turns that sheet into something a graph can check. Feed it frames, get back a boolean, a score, the worst frame's numbers, and the image itself so you can keep chaining. That's a delivery QC gate you can put in front of a render queue instead of a person with a waveform.

Its scope is deliberately narrow, and the node says so: picture only. There's no audio input, so no loudness check. If you need loudness, that's a different tool.

The two modes are the trick

mode has Preset and Guard, and the workflow the author intends uses both.

In Preset mode the image is ignored and the node emits a policy instead: data1 is a policy JSON and data2 is a human-readable description of it. You pick from Broadcast SDR, Cinema HDR (P3-PQ), OTT HDR10, Social Media or Custom, and Custom uses the five custom_* fields (max peak nits, max clipping, max black crush, max saturation). Optionally you can point policy_file at a JSON file and it replaces the preset; a failed load logs a warning and falls back rather than killing the run.

In Guard mode you wire that data1 into the policy input and the node checks the image against it. When policy is connected it replaces every max_* and require_metadata value - so the preset path and the guard path can't disagree about what the spec is. If you don't connect a policy, the node falls back to the inline max_clipping, max_black_crush, max_saturation, max_peak_nits and require_metadata widgets, which is fine for a quick one-off.

The defaults in the box are worth reading because they're honest about what they measure. max_clipping 0.01 means at most 1% of pixels with luma above 0.99. max_black_crush 0.05 means at most 5% below 0.01. max_saturation at 1.0 is effectively "only fail on negative pixel values", because mean (max - min) / max saturation can't exceed 1 on legal values - so if you want a real saturation gate, set it lower on purpose.

The check that actually bites

signal is how pixel values are read as light for the peak test, and getting it wrong makes the node lie to you. Display SDR (1.0 = 100 nits) for BT.1886 white; Scene-linear (1.0 = 203 nits) for Radiance's HDR convention (BT.2408); PQ (ST 2084), absolute code values; HLG (BT.2100, 1000-nit display), HLG code values on the reference display. Clipping, black crush and luma are measured on the values as they are, but peak nits depend entirely on this one setting - pass a PQ master with signal still on Display SDR and your 1000-nit highlights read as something like 10 nits. Nothing errors. You just get a pass on a spec violation.

metadata_present takes the metadata your deliverable actually carries, as comma-separated keys or key=value pairs (only the keys are checked), and require_metadata lists the keys that must appear there - colorspace, eotf. That's how "did the deliverable ship tagged" becomes a machine check instead of a footnote in an email.

Outputs: image passes through, passed is the boolean, data1 and data2 are the policy JSON and its description (in Guard mode they describe the run), and score is an INT you can sort or threshold. The report is for the worst frame in the batch, which is the right call - you get the failure, not an average that hides it.

Install

Ships with Radiance; install via ComfyUI Manager (search Radiance, then restart and refresh) or:

cd ComfyUI/custom_nodes
git clone https://github.com/fxtd-studios/radiance.git
cd radiance
python -m pip install -r requirements.txt

No models, no GPU - it's arithmetic over the tensor. Fast enough to gate every delivery.

Where people get burned

• Wrong signal. Covered above; it's the single most common way to get a green light you didn't earn.
• Checking a display-referred image as scene-linear. The peak test will scream. Set signal to match what's actually on the wire.
• Expecting a per-frame list. You get the worst frame's numbers in the report. If you need per-frame detail, that's what the QC node's JSON report is for.
• Using it as an artistic judgement. It checks numbers, not whether the shot looks good. Nothing here replaces a human looking at the frame - it just stops the obviously broken ones from reaching them.