Rembrandt, rim and butterfly: named light setups
Named portrait setups encode a lamp position, not a mood. The five worth memorising, the extra clause each one needs, and where the label stops working.
Put "dramatic lighting" in an image prompt and watch what actually changes. Contrast climbs, saturation drops, the whole frame gets a bit more contrasty and a bit more teal. What does not move is the light. It is in the same place it was in the previous render, because "dramatic" is a description of how a picture feels once it is finished, and the model has no way to work backwards from a feeling to a lamp.
Named portrait setups are the opposite kind of word. "Rembrandt lighting" is not a mood, it is a coordinate: key source roughly forty-five degrees off the lens axis, raised above the subject's eyeline, far enough round that the shadow from the nose meets the shadow from the cheek and leaves one small lit triangle under the eye on the shadow side. That is a specific physical arrangement with a testable output, and it is the kind of instruction generation models actually execute.
A named setup is a coordinate, not a mood
The pattern that keeps showing up in vendor prompt guidance is to describe the physical setup rather than the feeling: source size, direction, colour temperature. Google's own Veo 3.1 prompting guide even puts cinematography at the front of its prompt skeleton rather than tacking it on at the end, which tells you how much weight that block carries.
Named setups are efficient because they compress two of those three variables into one token. Say "Rembrandt" and you have specified direction and elevation in a single word, leaving your remaining prompt budget for the things the label does not cover. Say "moody" and you have specified nothing, but you have nudged the grade — which is why prompts stuffed with atmosphere adjectives tend to come back desaturated and crushed rather than lit.
Two of the setups below behave most consistently in practice: Rembrandt lighting and rim light both move output on their own, in the same way backlit silhouette, contre-jour, practicals, volumetric and golden hour do. The rest of the named portrait family — butterfly, loop, split — are standard photographic terms that work on the same principle, but I would not bet a render on the label alone. Attach the geometry either way and you are covered whether or not the model recognises the name.
Setup per mood
| Setup | Where the key sits | What it reads as | The extra clause it needs |
|---|---|---|---|
| Butterfly (Paramount) | High, directly on the lens axis, angled down | Polished, symmetrical, cosmetic | Whether there is fill under the chin, and how deep the shadow under the nose goes |
| Loop | 30–45° off axis, slightly above eye level | Ordinary, competent, neutral | That the nose shadow stays separate from the cheek shadow |
| Rembrandt | ~45° off axis, higher than loop | Interior, painterly, weighted | The lit triangle on the shadow cheek, named explicitly |
| Split | 90° to one side, level with the eyes | Division, interrogation, threat | How far the shadow half falls — full black, or lifted |
| Rim / kicker | Behind the subject, off to one side, pointing back toward the lens | Separation from a dark background | What the front of the face is lit by, or that it is not lit at all |
The last row is the one that trips people. Rim light on its own is a silhouette, because you have specified the only source in the scene and put it behind the subject. If you want the edge glow and a readable face, you owe the prompt a second source. If you want the silhouette, say so with contre-jour or backlit silhouette and stop there.
The four things the label never sets
A named setup fixes angle and elevation. It leaves four properties open, and a model with an open property will fill it with the training-set average, which is soft, frontal and neutral.
Source size. This is what decides shadow hardness, and it is not implied by any of the names above. A large source close to the subject wraps and softens; a small source hardens the edge of every shadow. Say which: "a large softbox" or "a bare bulb" or "a hard, distant source".
Ratio. How dark the shadow side is allowed to go. Plain language beats numbers here, because the model has no light meter: "the shadow side falls to near-black", or "shadow side lifted by a white bounce just out of frame".
Colour temperature. Warm tungsten, cool daylight, or a split between them. This is also where you get to be interesting — a warm key with a cool rim is a two-word instruction with a large visual return.
Motivation. What is emitting the light, in the frame or just outside it. A named source ("a desk lamp at frame left", "the green glow of a monitor", "harsh fluorescent overheads") is worth more than any adjective, because it gives the model a reason for the light to fall the way you asked, and reasons survive better than instructions.
Assembled, the difference is stark. This does very little:
Rembrandt lighting, cinematic, dramatic, moody portrait
This does the work:
Rembrandt key from camera left, a large softbox raised above her eyeline,
warm tungsten; the lit triangle visible on the shadow cheek; shadow side
lifted slightly by a white bounce card out of frame right; background
falls off to deep grey
That prompt carries six concrete descriptors — side, source size, colour temperature, the triangle, the bounce, and the background falloff — and not one style adjective. The broader source-direction-quality formula behind this is covered in lighting prompts for AI video and images if you want the general case rather than the named shortcuts.
Where named setups stop working
Full-body and wide shots. Butterfly, loop, Rembrandt and split are all defined by what happens on a face. At a wide shot the face is a hundred pixels across and the label has nothing to bite on. Switch to describing the scene's lighting rather than the portrait's: source position relative to the room, what is in shadow, where the falloff lands.
Anything with two subjects. A named setup assumes one head at a known angle to one key. Put two people in frame at different angles and the label is ambiguous — the model will pick one interpretation and you will not know which until it renders. Specify the light relative to the room instead, and let the geometry sort itself out.
Across a sequence. A label alone is not a continuity mechanism. "Rembrandt" in shot one and "Rembrandt" in shot four can easily come back with the key on opposite sides, because the label constrains the angle magnitude and not the side. If it matters, write the side into every prompt: "key from camera left" in all four, every time.
Models that are being asked to relight rather than generate. If you already have the shot and only the light is wrong, a named setup in a fresh text-to-image prompt is the wrong tool — you will get a different person under nicer light. That is a relighting job, which holds the subject and moves the source, and it behaves differently enough to be worth treating as its own operation.
When you are testing which vocabulary a given model responds to, run the same prompt against several models in one go rather than sequentially — A/B testing models on one prompt is faster than forming a theory from one render each. Models with strong prompt adherence will honour a named setup as written; weaker ones average it back toward frontal fill, and you find that out in one batch instead of five.
FAQ
Do video models parse named setups the same way image models do?
Less reliably, and the reason is structural: a video prompt is carrying camera, subject, action and context as well, and lighting is competing with all of it. The workaround is to establish the light in a still first and drive the shot from that still, which turns the lighting question into a reference image question rather than a prompt question. That also fixes the side-drift problem, since the reference has an unambiguous key direction baked into its pixels.
Should I keep using "cinematic lighting" at all?
It is not harmful, it is just not doing what you think. It biases the grade toward higher contrast and lower saturation, which is occasionally what you want. What it never does is move a lamp. If you find it in a prompt, treat it as a spare token you could spend on source size or colour temperature instead.
How do I check whether the setup actually landed?
Look for the specific artefact each one promises. Rembrandt has a triangle on the shadow cheek — either it is there or the render ignored you. Split has a hard vertical division at the nose. Butterfly has a small symmetrical shadow directly under the nose. These are pass/fail checks you can run in two seconds, which is the real advantage of named setups over adjectives: they produce evidence you can grade.
What about the light in the background, not on the face?
Named setups say nothing about it, and that is the most common source of a portrait that looks subtly wrong. If the subject is lit from camera left and the background is evenly bright, nothing in the scene explains it. Add one clause about background falloff or one visible practical, and the frame stops looking sourceless.