One light, forty five degrees, slightly above the eyes. That is the whole trick, and it has a name.
One light, forty five degrees, slightly above the eyes. That is the whole trick, and it has a name.
Your model never saw a photon. It read what photographers typed.
You typed beautiful dramatic cinematic lighting and got a face lit from everywhere at once, evenly, flatly, like a passport photo with a glow filter. Then somebody in a thread told you to try Rembrandt lighting instead, you changed nothing else, and the shadow finally landed on one side of the face where it belonged. Same model. Same seed. Same everything except six letters.
That gap is not magic and it is not a secret keyword. It is a fact about what these models were trained on, and once you understand it, half the lighting problems in your renders stop being mysterious.
Here is the thing worth internalizing before any of the specific terms. A diffusion model has never encountered light. It has no renderer, no ray tracer, no concept of a source and a surface and a falloff curve. What it has is a very large pile of images, each paired with text somebody wrote about that image. Everything it appears to know about lighting is a statistical relationship between words and pixel patterns.
So ask who wrote the captions. Photographers, stock agencies, museum catalogs, film stills databases, portfolio sites, product listings. Those people have a working vocabulary, and it is not adjectives. It is setup names, equipment, times of day, and numbers.
Beautiful appears under everything, so it points nowhere. Rembrandt appears under one very specific pattern of shadow, so it points somewhere.Rembrandt lighting is not a mood. It is a geometry with a pass or fail test attached.
The key light sits at roughly a 45 degree angle to the subject and slightly above eye level. That single placement produces a shadow from the nose and the brow that falls across the far cheek, and the identifying feature is a small illuminated triangle sitting under the eye on the darker side of the face. Photographers call it the Rembrandt patch. There is a size test for it: the triangle should be no wider than the length of the eye directly above it, and no longer than the subject's nose. Miss on either dimension and it is a different pattern with a different name.
Fill goes on the opposite side at lower power, typically one to two stops under the key, enough to keep detail in the shadow without erasing the pattern that makes the shot work. For a classic low key portrait the ratio between the lit side and the shadow side runs around 3:1 or 4:1, and 4:1 is about two stops of separation.
Now look at that paragraph again and count how many caption writers have described a portrait using those exact words. Thousands. That is the entire reason the term steers a model.
| Term | What it pins down | Use it when |
|---|---|---|
Rembrandt lighting | Key at 45 degrees, above eye level, triangle of light on the shadow cheek | You want a face with real depth and one clear light direction |
split lighting | Light hard to one side, face divided down the middle, half lit half dark | You want severity, tension, or a hard character read |
butterfly lighting | Key directly in front and high, small shadow under the nose | Glamour, beauty, cheekbones, clean symmetry |
loop lighting | Key slightly off axis, small nose shadow looping down the cheek | The safe, flattering default for most faces |
rim lighting or backlight | Source behind the subject, bright edge separating them from the background | The subject is disappearing into a dark background |
practical lighting | Visible in frame sources doing the lighting: lamps, neon, screens, candles | You want the environment to explain the light |
Six terms. Each one collapses a paragraph of description into a word the training captions already contained. That is a better trade than any stack of adjectives you can build.
Colour temperature is measured in kelvin and the two anchors are worth knowing by heart. Tungsten light lands around 3200K and reads warm and orange. Daylight sits around 5600K, which is the reference for a high midday sun with no cloud cover, and reads neutral to cool.
Those numbers appear constantly in captions and technical notes, so they are usable directly. 3200K tungsten practicals is a more specific instruction than warm cosy lighting, and it carries the implication of the source type along with the colour. Mixing them on purpose is where it gets interesting, because a frame with a 3200K practical inside and 5600K daylight through a window is an extremely common real photograph and the model has seen thousands of them.
Time of day works the same way. Golden hour is the stretch shortly after sunrise or before sunset when the sun sits near the horizon, so its light travels through much more atmosphere, the blue scatters out, and what reaches the subject is orange and soft. Blue hour is the window when the sun is meaningfully below the horizon and only indirect light remains, which is why it produces no sharp shadows at all and typically lasts around twenty to thirty minutes. Those are two completely different pictures, and evening light asks for neither of them.
Light falls off with the inverse square of distance. Double the distance from source to subject and the subject receives a quarter of the light, which is two stops.
The practical consequence is why close, large sources look soft and distant, small sources look hard. A source held near the subject has a dramatic difference in intensity between the near cheek and the far cheek, which produces rapid falloff and shaped, wrapping light. Push the same source far back and the near and far cheek get almost the same amount, which flattens everything.
You cannot type the inverse square law into a prompt. What you can do is stop asking for contradictory things in the same breath. Soft wrapping light and harsh shadows describe opposite distances. When a render comes back muddy, half the time the cause is that the prompt requested two incompatible lighting conditions and the model averaged them, which is exactly what averaging looks like.
Time to argue with everything above, because it deserves it.
Naming a setup does not guarantee the geometry. The model is matching a caption distribution, not solving for shadow placement, so Rembrandt lighting gets you into the neighbourhood of that pattern and does not promise the triangle is the right size or on the correct side. If the exact pattern matters, you are going to end up reaching for structural control rather than words, and a depth or normal map through ControlNet will hold a light direction that a prompt only suggests.
The adjectives are also not worthless, which is the part people overcorrect on. Cinematic lighting does something. It pulls toward a cluster of film still aesthetics: lower key, more contrast, coloured separation, wider dynamic range. The complaint is not that it fails, it is that it is imprecise, and imprecise is fine when you do not care and expensive when you do. Use it as a seasoning on top of a named setup rather than as a substitute for one.
And there is a real risk in over specifying. Stack six lighting terms, a colour temperature, a time of day and a source list into one prompt and you will start fighting your own subject and pose tokens for attention, and something in the picture will quietly get worse. Two lighting decisions in a prompt is usually plenty: a setup and a colour, or a setup and a time of day. If you need a third, it probably belongs in a second pass rather than the first.
Try the smallest possible experiment tonight. Take a prompt you already like, delete every lighting adjective in it, and put back exactly one named setup and one kelvin value. Run the same seed. Most people find the picture got more specific and less pretty, in that order, which is the correct direction to be moving.