Treat AI Relight as a Photometric Hypothesis, Not a New Portrait

The 23 August DEV post was about splitting a prompt into change versus preserve. This one is narrower: light is a physical claim.

If a model "relights" a portrait, it is asserting a new key, a new fill, and a new set of shadows on the same objects. That is closer to a renderer than to a filter. It can be useful. It can also mint a different evening, a different room, or a different person while still looking like photography.

I tested the current MendTrue relight preset on the public synthetic night-interior sample — not a customer photo.

Input: dual sources are already visible — cool window on one side, warm lamp on the other. Synthetic site sample.

The extra instruction was a single photometric request:

Add a soft key light from camera left. Keep identity, pose, clothing, background geometry, crop and time of day.

Official site after sample for the relight preset (same synthetic scene as the input). Treat new highlights as a hypothesis until catchlights and wall geometry agree with the source. AI photometric edit, not a reshoot.

Invariants that are cheaper to name than to repair

A relight job fails in a different way than a prompt-contract job. The face can remain "the same" while the physics rot.

I write the invariants down before generating:

• identity and expression;
• hair volume and clothing edges;
• window, lamp, table, plant, picture frame — count and placement;
• camera height and crop;
• outdoor time of day (blue hour stays blue hour);
• number of practical lights in the room.

Then the change set is allowed to include only: intensity, softness, and direction of light on surfaces that already exist.

If the window grows, the lamp shade changes species, or a third sconce appears, the hypothesis is already false. Pretty skin will not save it.

Three tests that do not need a lab

These are the checks I actually run at 100% zoom.

1. Catchlights. Eyes pick up the lights that exist. If the source had a window and a lamp, a result with a studio octabox and no window reflection is a different shoot.
2. Shadow direction. Nose, collar, and table legs should agree with each other. Mixed shadows from imaginary overheads are a tell.
3. Contact with the room. The lamp still sits on the same table. The plant still occupies the same corner. Relight is not set dressing.

Only after those three do I look at whether the new key is actually softer or merely brighter.

Failure modes I would log like bugs

Symptom	Likely broken invariant	Next retry
Face looks younger or smoother	identity / skin texture	Relight only; forbid retouch words
Window view becomes a different city	background geometry	Name the window as frozen
Warm lamp turns into daylight	time of day	State "keep blue hour + practical lamp"
Crop recenters the subject	camera	Repeat with explicit crop lock
Noise vanishes into plastic	material	Relight ≠ denoise; do not combine yet

I generated denoise as a separate 1-credit run on the same source file, on purpose. Mixing both in one request would hide which tool invented the smoothness.

What I store with the result

For a lighting experiment that might be reused in a product shot or a profile crop:

• the untouched source;
• the exact relight sentence;
• the approved output under a filename that says relight;
• a note that outdoor twilight was an input fact, not a style LUT.

You can run the same photometric loop in the MendTrue relight workspace. Approve the room and the catchlights before you approve the mood.